We are compiling an embedded C++ application that is deployed in a shielded device in an environment bombarded with ionizing radiation. We are using GCC and cross-compiling for ARM. When deployed, our application generates some erroneous data and crashes more often than we would like. The hardware is designed for this environment, and our application has run on this platform for several years.
Are there changes we can make to our code, or compile-time improvements that can be made to identify/correct soft errors and memory-corruption caused by single event upsets? Have any other developers had success in reducing the harmful effects of soft errors on a long-running application?
Working for about 4-5 years with software/firmware development and environment testing of miniaturized satellites*, I would like to share my experience here.
*(miniaturized satellites are a lot more prone to single event upsets than bigger satellites due to its relatively small, limited sizes for its electronic components)
To be very concise and direct: there is no mechanism to recover from detectable, erroneous
situation by the software/firmware itself without, at least, one
copy of minimum working version of the software/firmware somewhere for recovery purpose - and with the hardware supporting the recovery (functional).
Now, this situation is normally handled both in the hardware and software level. Here, as you request, I will share what we can do in the software level.
...recovery purpose.... Provide ability to update/recompile/reflash your software/firmware in real environment. This is an almost must-have feature for any software/firmware in highly ionized environment. Without this, you could have redundant software/hardware as many as you want but at one point, they are all going to blow up. So, prepare this feature!
...minimum working version... Have responsive, multiple copies, minimum version of the software/firmware in your code. This is like Safe mode in Windows. Instead of having only one, fully functional version of your software, have multiple copies of the minimum version of your software/firmware. The minimum copy will usually having much less size than the full copy and almost always have only the following two or three features:
capable of listening to command from external system,
capable of updating the current software/firmware,
capable of monitoring the basic operation's housekeeping data.
...copy... somewhere... Have redundant software/firmware somewhere.
You could, with or without redundant hardware, try to have redundant software/firmware in your ARM uC. This is normally done by having two or more identical software/firmware in separate addresses which sending heartbeat to each other - but only one will be active at a time. If one or more software/firmware is known to be unresponsive, switch to the other software/firmware. The benefit of using this approach is we can have functional replacement immediately after an error occurs - without any contact with whatever external system/party who is responsible to detect and to repair the error (in satellite case, it is usually the Mission Control Centre (MCC)).
Strictly speaking, without redundant hardware, the disadvantage of doing this is you actually cannot eliminate all single point of failures. At the very least, you will still have one single point of failure, which is the switch itself (or often the beginning of the code). Nevertheless, for a device limited by size in a highly ionized environment (such as pico/femto satellites), the reduction of the single point of failures to one point without additional hardware will still be worth considering. Somemore, the piece of code for the switching would certainly be much less than the code for the whole program - significantly reducing the risk of getting Single Event in it.
But if you are not doing this, you should have at least one copy in your external system which can come in contact with the device and update the software/firmware (in the satellite case, it is again the mission control centre).
You could also have the copy in your permanent memory storage in your device which can be triggered to restore the running system's software/firmware
...detectable erroneous situation.. The error must be detectable, usually by the hardware error correction/detection circuit or by a small piece of code for error correction/detection. It is best to put such code small, multiple, and independent from the main software/firmware. Its main task is only for checking/correcting. If the hardware circuit/firmware is reliable (such as it is more radiation hardened than the rests - or having multiple circuits/logics), then you might consider making error-correction with it. But if it is not, it is better to make it as error-detection. The correction can be by external system/device. For the error correction, you could consider making use of a basic error correction algorithm like Hamming/Golay23, because they can be implemented more easily both in the circuit/software. But it ultimately depends on your team's capability. For error detection, normally CRC is used.
...hardware supporting the recovery Now, comes to the most difficult aspect on this issue. Ultimately, the recovery requires the hardware which is responsible for the recovery to be at least functional. If the hardware is permanently broken (normally happen after its Total ionizing dose reaches certain level), then there is (sadly) no way for the software to help in recovery. Thus, hardware is rightly the utmost importance concern for a device exposed to high radiation level (such as satellite).
In addition to the suggestion for above anticipating firmware's error due to single event upset, I would also like to suggest you to have:
Error detection and/or error correction algorithm in the inter-subsystem communication protocol. This is another almost must have in order to avoid incomplete/wrong signals received from other system
Filter in your ADC reading. Do not use the ADC reading directly. Filter it by median filter, mean filter, or any other filters - never trust single reading value. Sample more, not less - reasonably.
NASA has a paper on radiation-hardened software. It describes three main tasks:
Regular monitoring of memory for errors then scrubbing out those errors,
robust error recovery mechanisms, and
the ability to reconfigure if something no longer works.
Note that the memory scan rate should be frequent enough that multi-bit errors rarely occur, as most ECC memory can recover from single-bit errors, not multi-bit errors.
Robust error recovery includes control flow transfer (typically restarting a process at a point before the error), resource release, and data restoration.
Their main recommendation for data restoration is to avoid the need for it, through having intermediate data be treated as temporary, so that restarting before the error also rolls back the data to a reliable state. This sounds similar to the concept of "transactions" in databases.
They discuss techniques particularly suitable for object-oriented languages such as C++. For example
Software-based ECCs for contiguous memory objects
Programming by Contract: verifying preconditions and postconditions, then checking the object to verify it is still in a valid state.
And, it just so happens, NASA has used C++ for major projects such as the Mars Rover.
C++ class abstraction and encapsulation enabled rapid development and testing among multiple projects and developers.
They avoided certain C++ features that could create problems:
Exceptions
Templates
Iostream (no console)
Multiple inheritance
Operator overloading (other than new and delete)
Dynamic allocation (used a dedicated memory pool and placement new to avoid the possibility of system heap corruption).
Here are some thoughts and ideas:
Use ROM more creatively.
Store anything you can in ROM. Instead of calculating things, store look-up tables in ROM. (Make sure your compiler is outputting your look-up tables to the read-only section! Print out memory addresses at runtime to check!) Store your interrupt vector table in ROM. Of course, run some tests to see how reliable your ROM is compared to your RAM.
Use your best RAM for the stack.
SEUs in the stack are probably the most likely source of crashes, because it is where things like index variables, status variables, return addresses, and pointers of various sorts typically live.
Implement timer-tick and watchdog timer routines.
You can run a "sanity check" routine every timer tick, as well as a watchdog routine to handle the system locking up. Your main code could also periodically increment a counter to indicate progress, and the sanity-check routine could ensure this has occurred.
Implement error-correcting-codes in software.
You can add redundancy to your data to be able to detect and/or correct errors. This will add processing time, potentially leaving the processor exposed to radiation for a longer time, thus increasing the chance of errors, so you must consider the trade-off.
Remember the caches.
Check the sizes of your CPU caches. Data that you have accessed or modified recently will probably be within a cache. I believe you can disable at least some of the caches (at a big performance cost); you should try this to see how susceptible the caches are to SEUs. If the caches are hardier than RAM then you could regularly read and re-write critical data to make sure it stays in cache and bring RAM back into line.
Use page-fault handlers cleverly.
If you mark a memory page as not-present, the CPU will issue a page fault when you try to access it. You can create a page-fault handler that does some checking before servicing the read request. (PC operating systems use this to transparently load pages that have been swapped to disk.)
Use assembly language for critical things (which could be everything).
With assembly language, you know what is in registers and what is in RAM; you know what special RAM tables the CPU is using, and you can design things in a roundabout way to keep your risk down.
Use objdump to actually look at the generated assembly language, and work out how much code each of your routines takes up.
If you are using a big OS like Linux then you are asking for trouble; there is just so much complexity and so many things to go wrong.
Remember it is a game of probabilities.
A commenter said
Every routine you write to catch errors will be subject to failing itself from the same cause.
While this is true, the chances of errors in the (say) 100 bytes of code and data required for a check routine to function correctly is much smaller than the chance of errors elsewhere. If your ROM is pretty reliable and almost all the code/data is actually in ROM then your odds are even better.
Use redundant hardware.
Use 2 or more identical hardware setups with identical code. If the results differ, a reset should be triggered. With 3 or more devices you can use a "voting" system to try to identify which one has been compromised.
You may also be interested in the rich literature on the subject of algorithmic fault tolerance. This includes the old assignment: Write a sort that correctly sorts its input when a constant number of comparisons will fail (or, the slightly more evil version, when the asymptotic number of failed comparisons scales as log(n) for n comparisons).
A place to start reading is Huang and Abraham's 1984 paper "Algorithm-Based Fault Tolerance for Matrix Operations". Their idea is vaguely similar to homomorphic encrypted computation (but it is not really the same, since they are attempting error detection/correction at the operation level).
A more recent descendant of that paper is Bosilca, Delmas, Dongarra, and Langou's "Algorithm-based fault tolerance applied to high performance computing".
Writing code for radioactive environments is not really any different than writing code for any mission-critical application.
In addition to what has already been mentioned, here are some miscellaneous tips:
Use everyday "bread & butter" safety measures that should be present on any semi-professional embedded system: internal watchdog, internal low-voltage detect, internal clock monitor. These things shouldn't even need to be mentioned in the year 2016 and they are standard on pretty much every modern microcontroller.
If you have a safety and/or automotive-oriented MCU, it will have certain watchdog features, such as a given time window, inside which you need to refresh the watchdog. This is preferred if you have a mission-critical real-time system.
In general, use a MCU suitable for these kind of systems, and not some generic mainstream fluff you received in a packet of corn flakes. Almost every MCU manufacturer nowadays have specialized MCUs designed for safety applications (TI, Freescale, Renesas, ST, Infineon etc etc). These have lots of built-in safety features, including lock-step cores: meaning that there are 2 CPU cores executing the same code, and they must agree with each other.
IMPORTANT: You must ensure the integrity of internal MCU registers. All control & status registers of hardware peripherals that are writeable may be located in RAM memory, and are therefore vulnerable.
To protect yourself against register corruptions, preferably pick a microcontroller with built-in "write-once" features of registers. In addition, you need to store default values of all hardware registers in NVM and copy-down those values to your registers at regular intervals. You can ensure the integrity of important variables in the same manner.
Note: always use defensive programming. Meaning that you have to setup all registers in the MCU and not just the ones used by the application. You don't want some random hardware peripheral to suddenly wake up.
There are all kinds of methods to check for errors in RAM or NVM: checksums, "walking patterns", software ECC etc etc. The best solution nowadays is to not use any of these, but to use a MCU with built-in ECC and similar checks. Because doing this in software is complex, and the error check in itself could therefore introduce bugs and unexpected problems.
Use redundancy. You could store both volatile and non-volatile memory in two identical "mirror" segments, that must always be equivalent. Each segment could have a CRC checksum attached.
Avoid using external memories outside the MCU.
Implement a default interrupt service routine / default exception handler for all possible interrupts/exceptions. Even the ones you are not using. The default routine should do nothing except shutting off its own interrupt source.
Understand and embrace the concept of defensive programming. This means that your program needs to handle all possible cases, even those that cannot occur in theory. Examples.
High quality mission-critical firmware detects as many errors as possible, and then handles or ignores them in a safe manner.
Never write programs that rely on poorly-specified behavior. It is likely that such behavior might change drastically with unexpected hardware changes caused by radiation or EMI. The best way to ensure that your program is free from such crap is to use a coding standard like MISRA, together with a static analyser tool. This will also help with defensive programming and with weeding out bugs (why would you not want to detect bugs in any kind of application?).
IMPORTANT: Don't implement any reliance of the default values of static storage duration variables. That is, don't trust the default contents of the .data or .bss. There could be any amount of time between the point of initialization to the point where the variable is actually used, there could have been plenty of time for the RAM to get corrupted. Instead, write the program so that all such variables are set from NVM in run-time, just before the time when such a variable is used for the first time.
In practice this means that if a variable is declared at file scope or as static, you should never use = to initialize it (or you could, but it is pointless, because you cannot rely on the value anyhow). Always set it in run-time, just before use. If it is possible to repeatedly update such variables from NVM, then do so.
Similarly in C++, don't rely on constructors for static storage duration variables. Have the constructor(s) call a public "set-up" routine, which you can also call later on in run-time, straight from the caller application.
If possible, remove the "copy-down" start-up code that initializes .data and .bss (and calls C++ constructors) entirely, so that you get linker errors if you write code relying on such. Many compilers have the option to skip this, usually called "minimal/fast start-up" or similar.
This means that any external libraries have to be checked so that they don't contain any such reliance.
Implement and define a safe state for the program, to where you will revert in case of critical errors.
Implementing an error report/error log system is always helpful.
It may be possible to use C to write programs that behave robustly in such environments, but only if most forms of compiler optimization are disabled. Optimizing compilers are designed to replace many seemingly-redundant coding patterns with "more efficient" ones, and may have no clue that the reason the programmer is testing x==42 when the compiler knows there's no way x could possibly hold anything else is because the programmer wants to prevent the execution of certain code with x holding some other value--even in cases where the only way it could hold that value would be if the system received some kind of electrical glitch.
Declaring variables as volatile is often helpful, but may not be a panacea.
Of particular importance, note that safe coding often requires that dangerous
operations have hardware interlocks that require multiple steps to activate,
and that code be written using the pattern:
... code that checks system state
if (system_state_favors_activation)
{
prepare_for_activation();
... code that checks system state again
if (system_state_is_valid)
{
if (system_state_favors_activation)
trigger_activation();
}
else
perform_safety_shutdown_and_restart();
}
cancel_preparations();
If a compiler translates the code in relatively literal fashion, and if all
the checks for system state are repeated after the prepare_for_activation(),
the system may be robust against almost any plausible single glitch event,
even those which would arbitrarily corrupt the program counter and stack. If
a glitch occurs just after a call to prepare_for_activation(), that would imply
that activation would have been appropriate (since there's no other reason
prepare_for_activation() would have been called before the glitch). If the
glitch causes code to reach prepare_for_activation() inappropriately, but there
are no subsequent glitch events, there would be no way for code to subsequently
reach trigger_activation() without having passed through the validation check or calling cancel_preparations first [if the stack glitches, execution might proceed to a spot just before trigger_activation() after the context that called prepare_for_activation() returns, but the call to cancel_preparations() would have occurred between the calls to prepare_for_activation() and trigger_activation(), thus rendering the latter call harmless.
Such code may be safe in traditional C, but not with modern C compilers. Such compilers can be very dangerous in that sort of environment because aggressive they strive to only include code which will be relevant in situations that could come about via some well-defined mechanism and whose resulting consequences would also be well defined. Code whose purpose would be to detect and clean up after failures may, in some cases, end up making things worse. If the compiler determines that the attempted recovery would in some cases invoke undefined behavior, it may infer that the conditions that would necessitate such recovery in such cases cannot possibly occur, thus eliminating the code that would have checked for them.
This is an extremely broad subject. Basically, you can't really recover from memory corruption, but you can at least try to fail promptly. Here are a few techniques you could use:
checksum constant data. If you have any configuration data which stays constant for a long time (including hardware registers you have configured), compute its checksum on initialization and verify it periodically. When you see a mismatch, it's time to re-initialize or reset.
store variables with redundancy. If you have an important variable x, write its value in x1, x2 and x3 and read it as (x1 == x2) ? x2 : x3.
implement program flow monitoring. XOR a global flag with a unique value in important functions/branches called from the main loop. Running the program in a radiation-free environment with near-100% test coverage should give you the list of acceptable values of the flag at the end of the cycle. Reset if you see deviations.
monitor the stack pointer. In the beginning of the main loop, compare the stack pointer with its expected value. Reset on deviation.
What could help you is a watchdog. Watchdogs were used extensively in industrial computing in the 1980s. Hardware failures were much more common then - another answer also refers to that period.
A watchdog is a combined hardware/software feature. The hardware is a simple counter that counts down from a number (say 1023) to zero. TTL or other logic could be used.
The software has been designed as such that one routine monitors the correct operation of all essential systems. If this routine completes correctly = finds the computer running fine, it sets the counter back to 1023.
The overall design is so that under normal circumstances, the software prevents that the hardware counter will reach zero. In case the counter reaches zero, the hardware of the counter performs its one-and-only task and resets the entire system. From a counter perspective, zero equals 1024 and the counter continues counting down again.
This watchdog ensures that the attached computer is restarted in a many, many cases of failure. I must admit that I'm not familiar with hardware that is able to perform such a function on today's computers. Interfaces to external hardware are now a lot more complex than they used to be.
An inherent disadvantage of the watchdog is that the system is not available from the time it fails until the watchdog counter reaches zero + reboot time. While that time is generally much shorter than any external or human intervention, the supported equipment will need to be able to proceed without computer control for that timeframe.
This answer assumes you are concerned with having a system that works correctly, over and above having a system that is minimum cost or fast; most people playing with radioactive things value correctness / safety over speed / cost
Several people have suggested hardware changes you can make (fine - there's lots of good stuff here in answers already and I don't intend repeating all of it), and others have suggested redundancy (great in principle), but I don't think anyone has suggested how that redundancy might work in practice. How do you fail over? How do you know when something has 'gone wrong'? Many technologies work on the basis everything will work, and failure is thus a tricky thing to deal with. However, some distributed computing technologies designed for scale expect failure (after all with enough scale, failure of one node of many is inevitable with any MTBF for a single node); you can harness this for your environment.
Here are some ideas:
Ensure that your entire hardware is replicated n times (where n is greater than 2, and preferably odd), and that each hardware element can communicate with each other hardware element. Ethernet is one obvious way to do that, but there are many other far simpler routes that would give better protection (e.g. CAN). Minimise common components (even power supplies). This may mean sampling ADC inputs in multiple places for instance.
Ensure your application state is in a single place, e.g. in a finite state machine. This can be entirely RAM based, though does not preclude stable storage. It will thus be stored in several place.
Adopt a quorum protocol for changes of state. See RAFT for example. As you are working in C++, there are well known libraries for this. Changes to the FSM would only get made when a majority of nodes agree. Use a known good library for the protocol stack and the quorum protocol rather than rolling one yourself, or all your good work on redundancy will be wasted when the quorum protocol hangs up.
Ensure you checksum (e.g. CRC/SHA) your FSM, and store the CRC/SHA in the FSM itself (as well as transmitting in the message, and checksumming the messages themselves). Get the nodes to check their FSM regularly against these checksum, checksum incoming messages, and check their checksum matches the checksum of the quorum.
Build as many other internal checks into your system as possible, making nodes that detect their own failure reboot (this is better than carrying on half working provided you have enough nodes). Attempt to let them cleanly remove themselves from the quorum during rebooting in case they don't come up again. On reboot have them checksum the software image (and anything else they load) and do a full RAM test before reintroducing themselves to the quorum.
Use hardware to support you, but do so carefully. You can get ECC RAM, for instance, and regularly read/write through it to correct ECC errors (and panic if the error is uncorrectable). However (from memory) static RAM is far more tolerant of ionizing radiation than DRAM is in the first place, so it may be better to use static DRAM instead. See the first point under 'things I would not do' as well.
Let's say you have an 1% chance of failure of any given node within one day, and let's pretend you can make failures entirely independent. With 5 nodes, you'll need three to fail within one day, which is a .00001% chance. With more, well, you get the idea.
Things I would not do:
Underestimate the value of not having the problem to start off with. Unless weight is a concern, a large block of metal around your device is going to be a far cheaper and more reliable solution than a team of programmers can come up with. Ditto optical coupling of inputs of EMI is an issue, etc. Whatever, attempt when sourcing your components to source those rated best against ionizing radiation.
Roll your own algorithms. People have done this stuff before. Use their work. Fault tolerance and distributed algorithms are hard. Use other people's work where possible.
Use complicated compiler settings in the naive hope you detect more failures. If you are lucky, you may detect more failures. More likely, you will use a code-path within the compiler which has been less tested, particularly if you rolled it yourself.
Use techniques which are untested in your environment. Most people writing high availability software have to simulate failure modes to check their HA works correctly, and miss many failure modes as a result. You are in the 'fortunate' position of having frequent failures on demand. So test each technique, and ensure its application actual improves MTBF by an amount that exceeds the complexity to introduce it (with complexity comes bugs). Especially apply this to my advice re quorum algorithms etc.
Since you specifically ask for software solutions, and you are using C++, why not use operator overloading to make your own, safe datatypes? For example:
Instead of using uint32_t (and double, int64_t etc), make your own SAFE_uint32_t which contains a multiple (minimum of 3) of uint32_t. Overload all of the operations you want (* + - / << >> = == != etc) to perform, and make the overloaded operations perform independently on each internal value, ie don't do it once and copy the result. Both before and after, check that all of the internal values match. If values don't match, you can update the wrong one to the value with the most common one. If there is no most-common value, you can safely notify that there is an error.
This way it doesn't matter if corruption occurs in the ALU, registers, RAM, or on a bus, you will still have multiple attempts and a very good chance of catching errors. Note however though that this only works for the variables you can replace - your stack pointer for example will still be susceptible.
A side story: I ran into a similar issue, also on an old ARM chip. It turned out to be a toolchain which used an old version of GCC that, together with the specific chip we used, triggered a bug in certain edge cases that would (sometimes) corrupt values being passed into functions. Make sure your device doesn't have any problems before blaming it on radio-activity, and yes, sometimes it is a compiler bug =)
Disclaimer: I'm not a radioactivity professional nor worked for this kind of application. But I worked on soft errors and redundancy for long term archival of critical data, which is somewhat linked (same problem, different goals).
The main problem with radioactivity in my opinion is that radioactivity can switch bits, thus radioactivity can/will tamper any digital memory. These errors are usually called soft errors, bit rot, etc.
The question is then: how to compute reliably when your memory is unreliable?
To significantly reduce the rate of soft errors (at the expense of computational overhead since it will mostly be software-based solutions), you can either:
rely on the good old redundancy scheme, and more specifically the more efficient error correcting codes (same purpose, but cleverer algorithms so that you can recover more bits with less redundancy). This is sometimes (wrongly) also called checksumming. With this kind of solution, you will have to store the full state of your program at any moment in a master variable/class (or a struct?), compute an ECC, and check that the ECC is correct before doing anything, and if not, repair the fields. This solution however does not guarantee that your software can work (simply that it will work correctly when it can, or stops working if not, because ECC can tell you if something is wrong, and in this case you can stop your software so that you don't get fake results).
or you can use resilient algorithmic data structures, which guarantee, up to a some bound, that your program will still give correct results even in the presence of soft errors. These algorithms can be seen as a mix of common algorithmic structures with ECC schemes natively mixed in, but this is much more resilient than that, because the resiliency scheme is tightly bounded to the structure, so that you don't need to encode additional procedures to check the ECC, and usually they are a lot faster. These structures provide a way to ensure that your program will work under any condition, up to the theoretical bound of soft errors. You can also mix these resilient structures with the redundancy/ECC scheme for additional security (or encode your most important data structures as resilient, and the rest, the expendable data that you can recompute from the main data structures, as normal data structures with a bit of ECC or a parity check which is very fast to compute).
If you are interested in resilient data structures (which is a recent, but exciting, new field in algorithmics and redundancy engineering), I advise you to read the following documents:
Resilient algorithms data structures intro by Giuseppe F.Italiano, Universita di Roma "Tor Vergata"
Christiano, P., Demaine, E. D., & Kishore, S. (2011). Lossless fault-tolerant data structures with additive overhead. In Algorithms and Data Structures (pp. 243-254). Springer Berlin Heidelberg.
Ferraro-Petrillo, U., Grandoni, F., & Italiano, G. F. (2013). Data structures resilient to memory faults: an experimental study of dictionaries. Journal of Experimental Algorithmics (JEA), 18, 1-6.
Italiano, G. F. (2010). Resilient algorithms and data structures. In Algorithms and Complexity (pp. 13-24). Springer Berlin Heidelberg.
If you are interested in knowing more about the field of resilient data structures, you can checkout the works of Giuseppe F. Italiano (and work your way through the refs) and the Faulty-RAM model (introduced in Finocchi et al. 2005; Finocchi and Italiano 2008).
/EDIT: I illustrated the prevention/recovery from soft-errors mainly for RAM memory and data storage, but I didn't talk about computation (CPU) errors. Other answers already pointed at using atomic transactions like in databases, so I will propose another, simpler scheme: redundancy and majority vote.
The idea is that you simply do x times the same computation for each computation you need to do, and store the result in x different variables (with x >= 3). You can then compare your x variables:
if they all agree, then there's no computation error at all.
if they disagree, then you can use a majority vote to get the correct value, and since this means the computation was partially corrupted, you can also trigger a system/program state scan to check that the rest is ok.
if the majority vote cannot determine a winner (all x values are different), then it's a perfect signal for you to trigger the failsafe procedure (reboot, raise an alert to user, etc.).
This redundancy scheme is very fast compared to ECC (practically O(1)) and it provides you with a clear signal when you need to failsafe. The majority vote is also (almost) guaranteed to never produce corrupted output and also to recover from minor computation errors, because the probability that x computations give the same output is infinitesimal (because there is a huge amount of possible outputs, it's almost impossible to randomly get 3 times the same, even less chances if x > 3).
So with majority vote you are safe from corrupted output, and with redundancy x == 3, you can recover 1 error (with x == 4 it will be 2 errors recoverable, etc. -- the exact equation is nb_error_recoverable == (x-2) where x is the number of calculation repetitions because you need at least 2 agreeing calculations to recover using the majority vote).
The drawback is that you need to compute x times instead of once, so you have an additional computation cost, but's linear complexity so asymptotically you don't lose much for the benefits you gain. A fast way to do a majority vote is to compute the mode on an array, but you can also use a median filter.
Also, if you want to make extra sure the calculations are conducted correctly, if you can make your own hardware you can construct your device with x CPUs, and wire the system so that calculations are automatically duplicated across the x CPUs with a majority vote done mechanically at the end (using AND/OR gates for example). This is often implemented in airplanes and mission-critical devices (see triple modular redundancy). This way, you would not have any computational overhead (since the additional calculations will be done in parallel), and you have another layer of protection from soft errors (since the calculation duplication and majority vote will be managed directly by the hardware and not by software -- which can more easily get corrupted since a program is simply bits stored in memory...).
One point no-one seems to have mentioned. You say you're developing in GCC and cross-compiling onto ARM. How do you know that you don't have code which makes assumptions about free RAM, integer size, pointer size, how long it takes to do a certain operation, how long the system will run for continuously, or various stuff like that? This is a very common problem.
The answer is usually automated unit testing. Write test harnesses which exercise the code on the development system, then run the same test harnesses on the target system. Look for differences!
Also check for errata on your embedded device. You may find there's something about "don't do this because it'll crash, so enable that compiler option and the compiler will work around it".
In short, your most likely source of crashes is bugs in your code. Until you've made pretty damn sure this isn't the case, don't worry (yet) about more esoteric failure modes.
You want 3+ slave machines with a master outside the radiation environment. All I/O passes through the master which contains a vote and/or retry mechanism. The slaves must have a hardware watchdog each and the call to bump them should be surrounded by CRCs or the like to reduce the probability of involuntary bumping. Bumping should be controlled by the master, so lost connection with master equals reboot within a few seconds.
One advantage of this solution is that you can use the same API to the master as to the slaves, so redundancy becomes a transparent feature.
Edit: From the comments I feel the need to clarify the "CRC idea." The possibilty of the slave bumping it's own watchdog is close to zero if you surround the bump with CRC or digest checks on random data from the master. That random data is only sent from master when the slave under scrutiny is aligned with the others. The random data and CRC/digest are immediately cleared after each bump. The master-slave bump frequency should be more than double the watchdog timeout. The data sent from the master is uniquely generated every time.
How about running many instances of your application. If crashes are due to random memory bit changes, chances are some of your app instances will make it through and produce accurate results. It's probably quite easy (for someone with statistical background) to calculate how many instances do you need given bit flop probability to achieve as tiny overall error as you wish.
What you ask is quite complex topic - not easily answerable. Other answers are ok, but they covered just a small part of all the things you need to do.
As seen in comments, it is not possible to fix hardware problems 100%, however it is possible with high probabily to reduce or catch them using various techniques.
If I was you, I would create the software of the highest Safety integrity level level (SIL-4). Get the IEC 61513 document (for the nuclear industry) and follow it.
Someone mentioned using slower chips to prevent ions from flipping bits as easily. In a similar fashion perhaps use a specialized cpu/ram that actually uses multiple bits to store a single bit. Thus providing a hardware fault tolerance because it would be very unlikely that all of the bits would get flipped. So 1 = 1111 but would need to get hit 4 times to actually flipped. (4 might be a bad number since if 2 bits get flipped its already ambiguous). So if you go with 8, you get 8 times less ram and some fraction slower access time but a much more reliable data representation. You could probably do this both on the software level with a specialized compiler(allocate x amount more space for everything) or language implementation (write wrappers for data structures that allocate things this way). Or specialized hardware that has the same logical structure but does this in the firmware.
Perhaps it would help to know does it mean for the hardware to be "designed for this environment". How does it correct and/or indicates the presence of SEU errors ?
At one space exploration related project, we had a custom MCU, which would raise an exception/interrupt on SEU errors, but with some delay, i.e. some cycles may pass/instructions be executed after the one insn which caused the SEU exception.
Particularly vulnerable was the data cache, so a handler would invalidate the offending cache line and restart the program. Only that, due to the imprecise nature of the exception, the sequence of insns headed by the exception raising insn may not be restartable.
We identified the hazardous (not restartable) sequences (like lw $3, 0x0($2), followed by an insn, which modifies $2 and is not data-dependent on $3), and I made modifications to GCC, so such sequences do not occur (e.g. as a last resort, separating the two insns by a nop).
Just something to consider ...
If your hardware fails then you can use mechanical storage to recover it. If your code base is small and have some physical space then you can use a mechanical data store.
There will be a surface of material which will not be affected by radiation. Multiple gears will be there. A mechanical reader will run on all the gears and will be flexible to move up and down. Down means it is 0 and up means it is 1. From 0 and 1 you can generate your code base.
Use a cyclic scheduler. This gives you the ability to add regular maintenance times to check the correctness of critical data. The problem most often encountered is corruption of the stack. If your software is cyclical you can reinitialize the stack between cycles. Do not reuse the stacks for interrupt calls, setup a separate stack of each important interrupt call.
Similar to the Watchdog concept is deadline timers. Start a hardware timer before calling a function. If the function does not return before the deadline timer interrupts then reload the stack and try again. If it still fails after 3/5 tries you need reload from ROM.
Split your software into parts and isolate these parts to use separate memory areas and execution times (Especially in a control environment). Example: signal acquisition, prepossessing data, main algorithm and result implementation/transmission. This means a failure in one part will not cause failures through the rest of the program. So while we are repairing the signal acquisition the rest of tasks continues on stale data.
Everything needs CRCs. If you execute out of RAM even your .text needs a CRC. Check the CRCs regularly if you using a cyclical scheduler. Some compilers (not GCC) can generate CRCs for each section and some processors have dedicated hardware to do CRC calculations, but I guess that would fall out side of the scope of your question. Checking CRCs also prompts the ECC controller on the memory to repair single bit errors before it becomes a problem.
Use watchdogs for bootup no just once operational. You need hardware help if your bootup ran into trouble.
Firstly, design your application around failure. Ensure that as part of normal flow operation, it expects to reset (depending on your application and the type of failure either soft or hard). This is hard to get perfect: critical operations that require some degree of transactionality may need to be checked and tweaked at an assembly level so that an interruption at a key point cannot result in inconsistent external commands.
Fail fast as soon as any unrecoverable memory corruption or control flow deviation is detected. Log failures if possible.
Secondly, where possible, correct corruption and continue. This means checksumming and fixing constant tables (and program code if you can) often; perhaps before each major operation or on a timed interrupt, and storing variables in structures that autocorrect (again before each major op or on a timed interrupt take a majority vote from 3 and correct if is a single deviation). Log corrections if possible.
Thirdly, test failure. Set up a repeatable test environment that flips bits in memory psuedo-randomly. This will allow you to replicate corruption situations and help design your application around them.
Given supercat's comments, the tendencies of modern compilers, and other things, I'd be tempted to go back to the ancient days and write the whole code in assembly and static memory allocations everywhere. For this kind of utter reliability I think assembly no longer incurs a large percentage difference of the cost.
Here are huge amount of replies, but I'll try to sum up my ideas about this.
Something crashes or does not work correctly could be result of your own mistakes - then it should be easily to fix when you locate the problem. But there is also possibility of hardware failures - and that's difficult if not impossible to fix in overall.
I would recommend first to try to catch the problematic situation by logging (stack, registers, function calls) - either by logging them somewhere into file, or transmitting them somehow directly ("oh no - I'm crashing").
Recovery from such error situation is either reboot (if software is still alive and kicking) or hardware reset (e.g. hw watchdogs). Easier to start from first one.
If problem is hardware related - then logging should help you to identify in which function call problem occurs and that can give you inside knowledge of what is not working and where.
Also if code is relatively complex - it makes sense to "divide and conquer" it - meaning you remove / disable some function calls where you suspect problem is - typically disabling half of code and enabling another half - you can get "does work" / "does not work" kind of decision after which you can focus into another half of code. (Where problem is)
If problem occurs after some time - then stack overflow can be suspected - then it's better to monitor stack point registers - if they constantly grows.
And if you manage to fully minimize your code until "hello world" kind of application - and it's still failing randomly - then hardware problems are expected - and there needs to be "hardware upgrade" - meaning invent such cpu / ram / ... -hardware combination which would tolerate radiation better.
Most important thing is probably how you get your logs back if machine fully stopped / resetted / does not work - probably first thing bootstap should do - is a head back home if problematic situation is entcovered.
If it's possible in your environment also to transmit a signal and receive response - you could try out to construct some sort of online remote debugging environment, but then you must have at least of communication media working and some processor/ some ram in working state. And by remote debugging I mean either GDB / gdb stub kind of approach or your own implementation of what you need to get back from your application (e.g. download log files, download call stack, download ram, restart)
I've really read a lot of great answers!
Here is my 2 cent: build a statistical model of the memory/register abnormality, by writing a software to check the memory or to perform frequent register comparisons. Further, create an emulator, in the style of a virtual machine where you can experiment with the issue. I guess if you vary junction size, clock frequency, vendor, casing, etc would observe a different behavior.
Even our desktop PC memory has a certain rate of failure, which however doesn't impair the day to day work.
I have a C++ program running on Linux in which a new thread is created to do some computationally expensive work independent of the main thread (The computational work completes by writing the results to files, which end up being very large). However, I'm getting relatively poor performance.
If I implement the program straightforward (without introducing other threads), it completes the task in roughly 2 hours. With the multi-threaded program it takes around 12 hours to do the same task (this was tested with only one thread spawned).
I've tried a couple of things, including pthread_setaffinity_np to set the thread to a single CPU (out of the 24 available on the server I'm using), as well as pthread_setschedparam to set the scheduling policy (I've only tried SCHED_BATCH). But the effects of these have so far been negligible.
Are there any general causes for this kind of problem?
EDIT: I've added some example code that I'm using, which is hopefully the most relevant parts. The function process_job() is what actually does the computational work, but it would be too much to include here. Basically, it reads in two files of data, and uses these to perform queries on an in-memory graph database, in which the results are written to two large files over a period of hours.
EDIT part 2: Just to clarify, the problem is not that I want to use threads to increase the performance of an algorithm I have. But rather, I want to run many instances of my algorithm simultaneously. Therefore, I expect the algorithm would run at a similar speed when put in a thread as it would if I didn't use multi-threads at all.
EDIT part 3: Thanks for the suggestions all. I'm currently doing some unit tests (seeing which parts are slowing down) as some have suggested. As the program takes a while to load and execute, it is taking time to see any results from the tests and therefore I apologize for late responses. I think the main point I wanted to clarify is possible reasons why threading could cause a program to run slowly. From what I gather from the comments, it simply shouldn't be. I'll post when I can find a reasonable resolution, thanks again.
(FINAL) EDIT part 4: It turns out that the problem was not related to threading after all. Describing it would be too cumbersome at this point (including the use of compiler optimization levels), but the ideas posted here were very useful and appreciated.
struct sched_param sched_param = {
sched_get_priority_min(SCHED_BATCH)
};
int set_thread_to_core(const long tid, const int &core_id) {
cpu_set_t mask;
CPU_ZERO(&mask);
CPU_SET(core_id, &mask);
return pthread_setaffinity_np(tid, sizeof(mask), &mask);
}
void *worker_thread(void *arg) {
job_data *temp = (job_data *)arg; // get the information for the task passed in
...
long tid = pthread_self();
int set_thread = set_thread_to_core(tid, slot_id); // assume slot_id is 1 (it is in the test case I run)
sched_get_priority_min(SCHED_BATCH);
pthread_setschedparam(tid, SCHED_BATCH, &sched_param);
int success = process_job(...); // this is where all the work actually happens
pthread_exit(NULL);
}
int main(int argc, char* argv[]) {
...
pthread_t temp;
pthread_create(&temp, NULL, worker_thread, (void *) &jobs[i]); // jobs is a vector of a class type containing information for the task
...
return 0;
}
If you have plenty of CPU cores, and have plenty of work to do, it should not take longer to run in multithreaded than single threaded mode - the actual CPU time may be a fraction longer, but the "wall-clock time" should be shorter. I'm pretty sure that your code has some sort of bottleneck where one thread is blocking the other.
This is because of one or more of these things - I'll list them first, then go into detail below:
Some lock in a thread is blocking the second thread from running.
Sharing of data between threads (either true or "false" sharing)
Cache thrashing.
Competition for some external resource causing thrashing and/or blocking.
Badly designed code in general...
Some lock in a thread is blocking the second thread from running.
If there is a thread that takes a lock, and another thread wants to use the resource that is locked by this thread, it will have to wait. This obviously means the thread isn't doing anything useful. Locks should be kept to a minimum by only taking the lock for a short period. Using some code to identify if locks are holding your code, such as:
while (!tryLock(some_some_lock))
{
tried_locking_failed[lock_id][thread_id]++;
}
total_locks[some_lock]++;
Printing some stats of the locks would help to identify where the locking is contentious - or you can try the old trick of "Press break in the debugger and see where you are" - if a thread is constantly waiting for some lock, then that's what's preventing progress...
Sharing of data between threads (either true or "false" sharing)
If two threads use [and update the value of it frequently] the same variable, then the two threads will have to swap "I've updated this" messages, and the CPU's have to fetch the data from the other CPU before it can continue with it's use of the variable. Since "data" is shared on a "per cache-line" level, and a cache-line is typically 32-bytes, something like:
int var[NUM_THREADS];
...
var[thread_id]++;
would classify as something called "false sharing" - the ACTUAL data updated is unique per CPU, but since the data is within the same 32-byte region, the cores will still have updated the same are of memory.
Cache thrashing.
If two threads do a lot of memory reading and writing, the cache of the CPU may be constantly throwing away good data to fill it with data for the other thread. There are some techniques available to ensure that two threads don't run in "lockstep" on which part of cache the CPU uses. If the data is 2^n (power of two) and fairly large (a multiple of the cache-size), it's a good idea to "add an offset" for each thread - for example 1KB or 2KB. That way, when the second thread reads the same distance into the data region, it will not overwrite exactly the same area of cache that the first thread is currently using.
Competition for some external resource causing thrashing and/or blocking.
If two threads are reading or writing from/to the hard-disk, network card, or some other shared resource, this can lead to one thread blocking another thread, which in turn means lower performance. It is also possible that the code detects different threads and does some extra flushing to ensure that data is written in the correct order or similar, before starting work with the other thread.
It is also possible that there are locks internally in the code that deals with the resource (user-mode library or kernel mode drivers) that block when more than one thread is using the same resource.
Generally bad design
This is a "catchall" for "lots of other things that can be wrong". If the result from one calculation in one thread is needed to progress the other, obviously, not a lot of work can be done in that thread.
Too small a work-unit, so all the time is spent starting and stopping the thread, and not enough work is being done. Say for example that you dole out small numbers to be "calculate if this is a prime" to each thread, one number at a time, it will probably take a lot longer to give the number to the thread than the calculation of "is this actually a prime-number" - the solution is to give a set of numbers (perhaps 10, 20, 32, 64 or such) to each thread, and then report back the result for the whole lot in one go.
There are plenty of other "bad design". Without understanding your code it's quite hard to say for sure.
It is entirely possible that your problem is none of the ones I've mentioned here, but most likely it is one of these. Hopefully this asnwer is helpful to identify the cause.
Read CPU Caches and Why You Care to understand why a naive port of an algorithm from one thread to multiple threads will more often than not result in greatly reduced performance and negative scalability. Algorithms that are specififcally designed for parallelism take care of overactive interlocked operations, false sharing and other causes of cache pollution.
Here are a few things you might wanna look into.
1°) Do you enter any critical section (locks, semaphores, etc.) between your worker thread and your main thread? (this should be the case if your queries modify the graph). If so, that could be one of the sources of the multithreading overhead : threads competing for a lock usually degrades performances.
2°) You're using a 24 cores machines, which I assume would be NUMA (Non-Uniform Memory Access). Since you set the threads affinities during your tests, you should pay close attention to the memory topology of your hardware. Looking at the files in /sys/devices/system/cpu/cpuX/ can help you with that (beware that cpu0 and cpu1 aren't necessarily close together, and thus does not necessarily share memory). Threads heavily using memory should use local memory (allocated in the same NUMA node as the core they're executing on).
3°) You are heavily using disk I/O. Which kind of I/O is that? if every thread perform every time some synchronous I/O, you might wanna consider asynchronous system calls, so that the OS stays in charge of scheduling those requests to the disk.
4°) Some caches issues have already been mentionned in other answers. From experience, false sharing can hurt performances as much as you're observing. My last recommendation (which should have been my first) is to use a profiler tool, such as Linux Perf, or OProfile. With such performance degradation you're experiencing, the cause will certainly appear quite clearly.
The other answers have all addressed the general guidelines that can cause your symptoms. I will give my own, hopefully not excessively redundant version. Then I will talk a bit about how you can get to the bottom of the problem with everything discussed in mind.
In general, there's a few reasons you'd expect multiple threads to perform better:
A piece of work is dependent on some resources (disk, memory, cache, etc.) while other pieces can proceed independently of these resources or said workload.
You have multiple CPU cores that can process your workload in parallel.
The main reasons, enumerated above, you'd expect multiple threads to perform less well are all based on resource contention:
Disk contention: already explained in detail and can be a possible issue, especially if you are writing small buffers at a time instead of batching
CPU time contention if the threads are scheduled onto the same core: probably not your issue if you're setting affinity. However, you should still double check
Cache thrashing: similarly probably not your problem if you have affinity, though this can be very expensive if it is your problem.
Shared memory: again talked about in detail and doesn't seem to be your issue, but it wouldn't hurt to audit the code to check it out.
NUMA: again talked about. If your worker thread is pinned to a different core, you will want to check whether the work it needs to access is local to the main core.
Ok so far not much new. It can be any or none of the above. The question is, for your case, how can you detect where the extra time is coming from. There's a few strategies:
Audit the code and look for obvious areas. Don't spend too much time doing this as it's generally unfruitful if you wrote the program to begin with.
Refactor the single threaded code and the multi-threaded code to isolate one process() function, then profile at key checkpoints to try to account for the difference. Then narrow it down.
Refactor the resource access into batches, then profile each batch on both the control and the experiment to account for the difference. Not only will this tell you which areas (disk access vs memory access vs spending time in some tight loop) you need to focus your efforts on, doing this refactor might even improve your running time overall. Example:
First copy the graph structure to thread-local memory (perform a straight-up copy in the single-threaded case)
Then perform the query
Then setup an asynchronous write to disk
Try to find a minimally reproducible workload with the same symptoms. This means changing your algorithm to do a subset of what it already does.
Make sure there's no other noise in the system that could've caused the difference (if some other user is running a similar system on the work core).
My own intuition for your case:
Your graph structure is not NUMA friendly for your worker core.
The kernel can actually scheduled your worker thread off the affinity core. This can happen if you don't have isolcpu on for the core you're pinning to.
I can't tell you what's wrong with your program because you haven't shared enough of it to do a detailed analysis.
What I can tell you is if this was my problem the first thing I would try is to run two profiler sessions on my application, one on the single threaded version and another on the dual thread configuration. The profiler report should give you a pretty good idea of where the extra time is going. Note that you may not need to profile the entire application run, depending on the problem the time difference may become obvious after you profile for a few seconds or minutes.
As far as profiler choices for Linux you may want to consider oprofile or as a second choice gprof.
If you find you need help interpreting the profiler output feel free to add that to your question.
It can be a right pain in the rear to track down why threads aren't working as planned. One can do so analytically, or one can use tool to show what's going on. I've had very good mileage out of ftrace, Linux's clone of Solaris's dtrace (which in turn is based on what VxWorks, Greenhill's Integrity OS and Mercury Computer Systems Inc have been doing for a looong time.)
In particular I found this page very useful: http://www.omappedia.com/wiki/Installing_and_Using_Ftrace, particularly this and this section. Don't worry about it being an OMAP orientated website; I've used it on X86 Linuxes just fine (though you may have to build a kernel to include it). Also remember that the GTKWave viewer is primarily intended for looking at log traces from VHDL developments, which is why it looks 'odd'. It's just that someone realised that it would be a usable viewer for sched_switch data too, and that saved them writing one.
Using the sched_switch tracer you can see when (but not necessarily why) your threads are running, and that might be enough to give you a clue. The 'why' can be revealed by careful examination of some of the other tracers.
If you are getting slowdown from using 1 thread, it is likely due to overhead from using thread safe library functions, or from thread setup. Creating a thread for each job will cause significant overhead, but probably not as much as you refer to.
In other words, it is probably some overhead from some thread safe library function.
The best thing to do, is to profile your code to find out where time is spent. If it is in a library call, try to find a replacement library or implement it yourself. If the bottleneck is thread creation/destruction try reusing threads, for instance using OpenMP tasks or std::async in C++11.
Some libraries are really nasty wrt thread safe overhead. For instance, many rand() implementations use a global lock, rather than using thread local prgn's. Such locking overhead is much larger than generating a number, and is hard to track without a profiler.
The slowdown could also stem from small changes you have made, for instance declaring variables volatile, which generally should not be necessary.
I suspect you're running on a machine with one single-core processor. This problem is not parallelizable on that kind of system. Your code is constantly using the processor, which has a fixed number of cycles to offer to it. It actually runs more slowly because the additional thread adds expensive context switching to the problem.
The only kinds of problems that parallelize well on a single-processor machine are those that allow one path of execution to run while another is blocked waiting for I/O, and situations (such as keeping a responsive GUI) where allowing one thread to get some processor time is more important than executing your code as quickly as possible.
If you only want to run many independent instances of your algorithm can you just submit multiple jobs (with different parameters, can be handled by a single script) to your cluster? That would eliminate the need to profile and debug your multithreaded program. I don't have much experience with multithreaded programming but if you use MPI or OpenMP then you'd have to write less code for the book keeping too. For example, if some common initialization routine is needed and the processes can run independently thereafter you can just do that by initializing in one thread and doing a broadcast. No need for maintaining locks and such.