In the simplified scenario where each object to be rendered is translated into a secondary command buffer and each of those command buffers bind a graphics pipeline initially: is a guaranteed no-op to bind the pipeline that was immediately bound before? Or the order of execution of the secondary command buffers is not guaranteed at all?
is a guaranteed no-op to bind the pipeline that was immediately bound before?
No. In fact, in the case you're outlining, you should assume precisely the opposite. Why?
Since each of your CBs is isolated from the others, the vkCmdBindPipeline function has no way to know what was bound beforehand. Remember: the state of a command buffer that has started recording is undefined. Which means that the command buffer building code cannot make any assumptions about any state which you did not set within this CB.
In order for the driver to implement the optimization you're talking about, it would have to, at vkCmdExecuteCommands time, introspect into each secondary command buffer and start ripping out anything that is duplicated across CB boundaries.
That might be viable if vkCmdExecuteCommands has to copy all of the commands out of secondary CBs into primary ones. But that would only be reasonable for systems where secondary CBs don't exist at a hardware level and thus have to be implemented by copying their commands into the primary CB. But even in this case, implementing such culling would make the command take longer to execute compared to simply copying some tokens into the primary CB's storage.
When dealing with a low-level API, do not assume that the driver is going to use information outside of its immediate purview to optimize your code. Especially when you have the tools for doing that optimization yourself.
This is (yet another) a reason why you should not give each individual object its own CB.
Or the order of execution of the secondary command buffers is not guaranteed at all?
The order of execution of commands is unchanged by their presence in CBs. However, the well-defined nature of the state these commands use is affected.
Outside of state inherited by secondary CBs, every secondary CB's state begins undefined. That's why you have to bind a pipeline for each one. Commands that rely on previously issued state only have well-defined behavior if that previously issued state is within the CB containing that command (or is inherited state).
Related
I would like to backup a running rocksdb-instance to a location on the same disk in a way that is safe, and without interrupting processing during the backup.
I have read:
Rocksdb Backup Instructions
Checkpoints Documentation
Documentation in rocksdb/utilities/{checkpoint.h,backupable_db.{h,cc}}
My question is whether the call to CreateNewBackupWithMetadata is marked as NOT threadsafe to express, that two concurrent calls to this function will have unsafe behavior, or to indicate that ANY concurrent call on the database will be unsafe. I have checked the implementation, which appears to be creating a checkpoint - which the second article claims are used for online backups of MyRocks -, but I am still unsure, what part of the call is not threadsafe.
I currently interpret this as, it is unsafe, because CreateBackup... calls DisableFileDeletions and later EnableFileDeletions, which, of course, if two overlapping calls are made, may cause trouble. Since the SST-files are immutable, I am not worried about them, but am unsure whether modifying the WAL through insertions can corrupt the backup. I would assume that triggering a flush on backup should prevent this, but I would like to be sure.
Any pointers or help are appreciated.
I ended up looking into the implementation way deeper, and here is what I found:
Recall a rocksdb database consists of Memtables, SSTs and a single WAL, which protects data in the Memtables against crashes.
When you call rocksdb::BackupEngine::CreateBackupWithMetadata, there is no lock taken internally, so this call can race, if two calls are active at the same time. Most notably this call does Disable/EnableFileDeletions, which, if called by one call, while another is still active spells doom for the other call.
The process of copying the files from the database to the backup is protected from modifications while the call is active by creating a rocksdb::Checkpoint, which, if flush_before_backup was set to true, will first flush the Memtables, thus clearing the active WAL.
Internally the call to CreateCustomCheckpoint calls DB::GetLiveFiles in db_filecheckpoint.cc. GetLiveFiles takes the global database lock (_mutex), optionally flushes the Memtables, and retrieves the list of SSTs. If a flush in GetLiveFiles happens while holding the global database-lock, the WAL must be empty at this time, which means the list should always contain the SST-files representing a complete and consistent database state from the time of the checkpoint. Since the SSTs are immutable, and since file deletion through compaction is turned off by the backup-call, you should always get a complete backup without holding writes on the database. However this, of course, means it is not possible to determine the exact last write/sequence number in the backup when concurrent updates happen - at least not without inspecting the backup after it has been created.
For the non-flushing version, there maybe WAL-files, which are retrieved in a different call than GetLiveFiles, with no lock held in between, i.e. these are not necessarily consistent, but I did not investigate further, since the non-flushing case was not applicable to my use.
I was trying to create a memory buffer in OpenCL with C++ binding. The sentence looks like
cl::Buffer buffer(context,CL_MEM_READ_ONLY,sizeof(float)*(100));
This sentence confuses me because it doesn't specify which device the memory is allocated on. In principle context contains all devices, including cpu and gpu, on the chosen platform. Is it true that the buffer is put in a common region shared by all the devices?
The spec does not define where the memory is. For the API user, it is "in the context".
If you have one device only, probably (99.99%) is going to be in the device. (In rare cases it may be in the host if the device does not have enough memory for the time being)
In case of many different devices, it will be in one of them at the creation. But it may move transparently to another device depending on the kernel launches.
This is the reason why the call clEnqueueMIgrateMemObjects (OpenCL 1.2 only) exists.
It allows the user to tell some hints to the API about where the memory will be needed, and prepare the copy in advance.
Here is the definition of what it does:
clEnqueueMIgrateMemObjects provides a mechanism for assigning which device an OpenCL memory object resides. A user may wish to have more explicit control over the location of their memory objects on creation. This could be used to:
Ensure that an object is allocated on a specific device prior to usage.
Preemptively migrate an object from one device to another.
Typically, memory objects are implicitly migrated to a device for which enqueued commands, using the memory object, are targeted
Looking at glUnmapBuffer at the moment, according to some docs it returns GLboolean to indicate success (GL_TRUE) or failure (GL_FALSE). Presumably if it fails it sets the GL error.
My question is under what circumstances you would need to check to see if it returned with GL_TRUE in release builds, assuming you're checking glGetError after every OpenGL call in debug builds?
Or rather, why would this function need to return a bool success/fail when other OpenGL functions do not? I'm looking for the organising principle here, i.e. can it sometimes return GL_FALSE without having set the gl error?
EDIT:
OK, I found the answer. Not sure whether to leave this up because it's interesting. On this page:
"Because of its low-level nature, these protections have to be
relaxed. Therefore, it is possible that, during the time a buffer is
mapped, some kind of corruption happens. If this occurs, calling
glUnmapBuffer​ will return GL_FALSE. At that point, the contents of
the buffer in question are considered undefined. It may have your
data, or it may have random garbage."
So as I'm specifying Windows 7 or later, I don't need to consider it (I think). Ironically, one of the reasons I switched out of Direct 3D into OpenGL was precisely this (with Direct 3D 9.0).
Because it's not a regular OpenGL error:
(...) This can occur for system-specific reasons that affect the availability of graphics memory, such as screen mode changes. In such situations, GL_FALSE is returned and the data store contents are undefined. An application must detect this rare condition and reinitialize the data store.
This special mechanism was undoubtfully introduced to separate such cases from typical user interaction errors caught by glGetError.
I have written a program (suppose X) in c++ which creates a data structure and then uses it continuously.
Now I would like to modify that data structure without aborting the previous program.
I tried 2 ways to accomplish this task :
In the same program X, first I created data structure and then tried to create a child process which starts accessing and using that data structure for some purpose. The parent process continues with its execution and asks the user for any modification like insertion, deletion, etc and takes input from console and subsequently modification is done. The problem here is, it doesn't modify the copy of data structure that the child process was using. Later on, I figured out this won't help because the child process is using its own copy of data structure and hence modifications done via parent process won't be reflected in it. But definitely, I didn't want this to happen. So I went for multithreading.
Instead of creating child process, I created an another thread which access that data structure and uses it and tried to take user input from console in different thread. Even,
this didn't work because of very fast switching between threads.
So, please help me to solve this issue. I want the modification to be reflected in the original data structure. Also I don't want the process (which is accessing and using it continuously) to wait for sometimes since it's time crucial.
First point: this is not a trivial problem. To handle it at all well, you need to design a system, not just a quick hack or two.
First of all, to support the dynamic changing, you'll almost certainly want to define the data structure in code in something like a DLL or .so, so you can load it dynamically.
Part of how to proceed will depend on whether you're talking about data that's stored strictly in memory, or whether it's more file oriented. In the latter case, some of the decisions will depend a bit on whether the new form of a data structure is larger than an old one (i.e., whether you can upgrade in place or no).
Let's start out simple, and assume you're only dealing with structures in memory. Each data item will be represented as an object. In addition to whatever's needed to access the data, each object will provide locking, and a way to build itself from an object of the previous version of the object (lazily -- i.e., on demand, not just in the ctor).
When you load the DLL/.so defining a new object type, you'll create a collection of those the same size as your current collection of existing objects. Each new object will be in the "lazy" state, where it's initialized, but hasn't really been created from the old object yet.
You'll then kick off a thread that walks makes the new collection known to the rest of the program, then walks through the collection of new objects, locking an old object, using it to create a new object, then destroying the old object and removing it from the old collection. It'll use a fairly short timeout when it tries to lock the old object (i.e., if an object is in use, it won't wait for it very long, just go on to the next. It'll iterate repeatedly until all the old objects have been updated and the collection of old objects is empty.
For data on disk, things can be just about the same, except your collections of objects provide access to the data on disk. You create two separate files, and copy data from one to the other, converting as needed.
Another possibility (especially if the data can be upgraded in place) is to use a single file, but embed a version number into each record. Read some raw data, check the version number, and use appropriate code to read/write it. If you're reading an old version number, read with the old code, convert to the new format, and write in the new format. If you don't have space to update in place, write the new record to the end of the file, and update the index to indicate the new position.
Your approach to concurrent access is similar to sharing a cake between a classroom full of blindfolded toddlers. It's no surprise that you end up with a sticky mess. Each toddler will either have to wait their turn to dig in or know exactly which part of the cake she alone can touch.
Translating to code, the former means having a lock or mutex that controls access to a data structure so that only one thread can modify it at any time.
The latter can be done by having a data structure that is modified in place by threads that each know exactly which parts of the data structure they can update, e.g. by passing a struct with details on which range to update, effectively splitting up the data beforehand. These should not overlap and iterators should not be invalidated (e.g. by resizing), which may not be possible for a given problem.
There are many many algorithms for handling resource competition, so this is grossly simplified. Distributed computing is a significant field of computer science dedicated to these kinds problems; study the problem (you didn't give details) and don't expect magic.
In Java each object has a synchronisation monitor. So i guess the implementation is pretty condensed in term of memory usage and hopefully fast as well.
When porting this to C++ what whould be the best implementation for it. I think that there must be something better then "pthread_mutex_init" or is the object overhead in java really so high?
Edit: i just checked that pthread_mutex_t on Linux i386 is 24 bytes large. Thats huge if i have to reserve this space for each object.
In a sense it's worse than pthread_mutex_init, actually. Because of Java's wait/notify you kind of need a paired mutex and condition variable to implement a monitor.
In practice, when implementing a JVM you hunt down and apply every single platform-specific optimisation in the book, and then invent some new ones, to make monitors as fast as possible. If you can't do a really fiendish job of that, you definitely aren't up to optimising garbage collection ;-)
One observation is that not every object needs to have its own monitor. An object which isn't currently synchronised doesn't need one. So the JVM can create a pool of monitors, and each object could just have a pointer field, which is filled in when a thread actually wants to synchronise on the object (with a platform-specific atomic compare and swap operation, for instance). So the cost of monitor initialisation doesn't have to add to the cost of object creation. Assuming the memory is pre-cleared, object creation can be: decrement a pointer (plus some kind of bounds check, with a predicted-false branch to the code that runs gc and so on); fill in the type; call the most derived constructor. I think you can arrange for the constructor of Object to do nothing, but obviously a lot depends on the implementation.
In practice, the average Java application isn't synchronising on very many objects at any one time, so monitor pools are potentially a huge optimisation in time and memory.
The Sun Hotspot JVM implements thin locks using compare and swap. If an object is locked, then the waiting thread wait on the monitor of thread which locked the object. This means you only need one heavy lock per thread.
I'm not sure how Java does it, but .NET doesn't keep the mutex (or analog - the structure that holds it is called "syncblk" there) directly in the object. Rather, it has a global table of syncblks, and object references its syncblk by index in that table. Furthermore, objects don't get a syncblk as soon as they're created - instead, it's created on demand on the first lock.
I assume (note, I do not know how it actually does that!) that it uses atomic compare-and-exchange to associate the object and its syncblk in a thread-safe way:
Check the hidden syncblk_index field of our object for 0. If it's not 0, lock it and proceed, otherwise...
Create a new syncblk in global table, get the index for it (global locks are acquired/released here as needed).
Compare-and-exchange to write it into object itself.
If previous value was 0 (assume that 0 is not a valid index, and is the initial value for the hidden syncblk_index field of our objects), our syncblk creation was not contested. Lock on it and proceed.
If previous value was not 0, then someone else had already created a syncblk and associated it with the object while we were creating ours, and we have the index of that syncblk now. Dispose the one we've just created, and lock on the one that we've obtained.
Thus the overhead per-object is 4 bytes (assuming 32-bit indices into syncblk table) in best case, but larger for objects which actually have been locked. If you only rarely lock on your objects, then this scheme looks like a good way to cut down on resource usage. But if you need to lock on most or all your objects eventually, storing a mutex directly within the object might be faster.
Surely you don't need such a monitor for every object!
When porting from Java to C++, it strikes me as a bad idea to just copy everything blindly. The best structure for Java is not the same as the best for C++, not least because Java has garbage collection and C++ doesn't.
Add a monitor to only those objects that really need it. If only some instances of a type need synchronization then it's not that hard to create a wrapper class that contains the mutex (and possibly condition variable) necessary for synchronization. As others have already said, an alternative is to use a pool of synchronization objects with some means of choosing one for each object, such as using a hash of the object address to index the array.
I'd use the boost thread library or the new C++0x standard thread library for portability rather than relying on platform specifics at each turn. Boost.Thread supports Linux, MacOSX, win32, Solaris, HP-UX and others. My implementation of the C++0x thread library currently only supports Windows and Linux, but other implementations will become available in due course.