How can I fully utilize each of my EC2 cores?
I'm using a c4.4xlarge AWS Ubuntu EC2 instance and TensorFlow to build a large convoluted neural network. nproc says that my EC2 instance has 16 cores. When I run my convnet training code, the top utility says that I'm only using 400% CPU. I was expecting it to use 1600% CPU because of the 16 cores. The AWS EC2 monitoring tab confirms that I'm only using 25% of my CPU capacity. This is a huge network, and on my new Mac Pro it consumes about 600% CPU and takes a few hours to build, so I don't think the reason is because my network is too small.
I believe the line below ultimately determines CPU usage:
sess = tf.InteractiveSession(config=tf.ConfigProto())
I admit I don't fully understand the relationship between threads and cores, but I tried increasing the number of cores. It had the same effect as the line above: still 400% CPU.
NUM_THREADS = 16
sess = tf.InteractiveSession(config=tf.ConfigProto(intra_op_parallelism_threads=NUM_THREADS))
EDIT:
htop shows that shows that I am actually using all 16 of my EC2 cores, but each core is only at about 25%
top shows that my total CPU % is around 400%, but occasionally it will shoot up to 1300% and then almost immediately go back down to ~400%. This makes me think there could be a deadlock problem
Several things you can try:
Increase the number of threads
You already tried changing the intra_op_parallelism_threads. Depending on your network it can also make sense to increase the inter_op_parallelism_threads. From the doc:
inter_op_parallelism_threads:
Nodes that perform blocking operations are enqueued on a pool of
inter_op_parallelism_threads available in each process. 0 means the
system picks an appropriate number.
intra_op_parallelism_threads:
The execution of an individual op (for
some op types) can be parallelized on a pool of
intra_op_parallelism_threads. 0 means the system picks an appropriate
number.
(Side note: The values from the configuration file referenced above are not the actual default values tensorflow uses but just example values. You can see the actual default configuration by manually inspecting the object returned by tf.ConfigProto().)
Tensorflow uses 0 for the above options meaning it tries to choose appropriate values itself. I don't think tensorflow picked poor values that caused your problem but you can try out different values for the above option to be on the safe side.
Extract traces to see how well your code parallelizes
Have a look at
tensorflow code optimization strategy
It gives you something like this. In this picture you can see that the actual computation happens on far fewer threads than available. This could also be the case for your network. I marked potential synchronization points. There you can see that all threads are active for a short moment which potentially is the reason for the sporadic peaks in CPU utilization that you experience.
Miscellaneous
Make sure you are not running out of memory (htop)
Make sure you are not doing a lot of I/O or something similar
Related
I want to see programmatically how much GPU time a part of my application consumes on macOS and iOS. On OpenGL and D3D I can use GPU timer query objects. I searched and couldn't find anything similar for Metal. How do I measure GPU time on Metal without using Instruments etc. I'm using Objective-C.
There are a couple of problems with this method:
1) You really want to know what is the GPU side latency within a command buffer most of the time, not round trip to CPU. This is better measured as the time difference between running 20 instances of the shader and 10 instances of the shader. However, that approach can add noise since the error is the sum of the errors associated with the two measurements.
2) Waiting for completion causes the GPU to clock down when it stops executing. When it starts back up again, the clock is in a low power state and may take quite a while to come up again, skewing your results. This can be a serious problem and may understate your performance in benchmark vs. actual by a factor of two or more.
3) if you start the clock on scheduled and stop on completed, but the GPU is busy running other work, then your elapsed time includes time spent on the other workload. If the GPU is not busy, then you get the clock down problems described in (2).
This problem is considerably harder to do right than most benchmarking cases I've worked with, and I have done a lot of performance measurement.
The best way to measure these things is to use on device performance monitor counters, as it is a direct measure of what is going on, using the machine's own notion of time. I favor ones that report cycles over wall clock time because that tends to weed out clock slewing, but there is not universal agreement about that. (Not all parts of the hardware run at the same frequency, etc.) I would look to the developer tools for methods to measure based on PMCs and if you don't find them, ask for them.
You can add scheduled and completed handler blocks to a command buffer. You can take timestamps in each and compare. There's some latency, since the blocks are executed on the CPU, but it should get you close.
With Metal 2.1, Metal now provides "events", which are more like fences in other APIs. (The name MTLFence was already used for synchronizing shared heap stuff.) In particular, with MTLSharedEvent, you can encode commands to modify the event's value at particular points in the command buffer(s). Then, you can either way for the event to have that value or ask for a block to be executed asynchronously when the event reaches a target value.
That still has problems with latency, etc. (as Ian Ollmann described), but is more fine grained than command buffer scheduling and completion. In particular, as Klaas mentions in a comment, a command buffer being scheduled does not indicate that it has started executing. You could put commands to set an event's value at the beginning and (with a different value) at the end of a sequence of commands, and those would only notify at actual execution time.
Finally, on iOS 10.3+ but not macOS, MTLCommandBuffer has two properties, GPUStartTime and GPUEndTime, with which you can determine how much time a command buffer took to execute on the GPU. This should not be subject to latency in the same way as the other techniques.
As an addition to Ken's comment above, GPUStartTime and GPUEndTime is now available on macOS too (10.15+):
https://developer.apple.com/documentation/metal/mtlcommandbuffer/1639926-gpuendtime?language=objc
I have a certain program that recieves input and returnes an output with a run-time of about 2 seconds,
Now, i want to run this program online on a server that can handle multiple connections (lets say up to 100k),
on each client-server session the program will launch,
the client will hand the server the program's input and will wait for the program to end to recieve the server's respond (program's output),
Lets say the server's host is a very powerful machine - e.g 16 cores,
Can this work or it is to much runtime for each client?
What is the maximum runtime this kind of program can have?
I'm posting this as an answer because it's too large to place as a comment.
Can this work? It depends. It depends because there are a lot of variables in this problem. Let's look at some of them:
you say it takes 2 seconds to compute a result. Where and how are those seconds spent? Is this pure computation or are you accessing a database, or the file system? Is this CPU bound or I/O bound? If you run computations for the full 2 seconds then you are consuming CPU which means that you can simultaneously serve only 16 clients, one per core. Are you hitting a database? Is this on the powerful server or on some other machine? If the database is the bottleneck than move this to the powerful server and have SSD drives on it.
can you improve processing for one client? What's more efficient, doing the processing on one core or spread it across all the cores? If you can parallelize, can you limit thread contention?
is CPU all you need? How about memory? Any backend service you access? Are those over the network? Do you have enough bandwidth?
related to memory, what language/platform are you using? Does it have a garbage collector? Do you generate a lot of object to compute a result? Does the GC kick in and pauses your application so it cleans up and compacts the memory? Do you allocate enough memory for the application to run?
can you cache responses and serve them to other clients or are responses custom to each client? Can you precompute the results and then just serve them to clients or can't you predict the inputs?
Have you tried running some performance tests and profile the application to see where hotspots might show up? Can you do something about them?
have you any imposed performance criteria? How many clients do you want to support simultaneously? Is 2 seconds too much? Can clients live with more? How much more? How many seconds does it mean an unacceptable response time?
do you need a big server to run this setup or smaller ones work better (i.e. scale horizontally instead of vertically)?
etc
Nobody can answer this for you. You have to do an analysis of your application, run some tests, profile it, optimize it, then repeat until you are satisfied with the results.
I implemented a simple http server link, but the result of the test (ab -n 10000 -c 100 http://localhost:8080/status) is very bad (look through the test.png in the previous link)
I don't understand why it doesn't work correctly with multiple threads.
I believe that, by default, Netty's default thread pool is configured with as many threads as there are cores on the machine. The idea being to handle requests asynchronously and non-blocking (where possible).
Your /status test includes a database transaction which blocks because of the intrinsic design of database drivers etc. So your performance - at high level - is essentially a result of:-
a.) you are running a pretty hefty test of 10,000 requests attempting to run 100 requests in parallel
b.) you are calling into a database for each request so this is will not be quick (relatively speaking compared to some non-blocking I/O operation)
A couple of questions/considerations for you:-
Machine Spec.?
What is the spec. of the machine you are running your application and test on?
How many cores?
If you only have 8 cores available then you will only have 8 threads running in parallel at any time. That means those batches of 100 requests per time will be queueing up
Consider what is running on the machine during the test
It sound like you are running the application AND Apache Bench on the same machine so be aware that both your application and the testing tool will both be contending for those cores (this is in addition to any background processes going on also contending for those cores - such as the OS)
What will the load be?
Predicting load is difficult right. If you do think you are likely to have 100 requests into the database at any one time then you may need to think about:-
a. your production environment may need a couple of instance to handle the load
b. try changing the config. of Netty's default thread pool to increase the number of threads
c. think about your application architecture - can you cache any of those results instead of going to the database for each request
May be linked to the usage of Database access (synchronous task) within one of your handler (at least in your TrafficShappingHandler) ?
You might need to "make async" your database calls (other threads in a producer/consumer way for instance)...
If something else, I do not have enough information...
I have been trying to load test my API server using Locust.io on EC2 compute optimized instances. It provides an easy-to-configure option for setting the consecutive request wait time and number of concurrent users. In theory, rps = wait time X #_users. However while testing, this rule breaks down for very low thresholds of #_users (in my experiment, around 1200 users). The variables hatch_rate, #_of_slaves, including in a distributed test setting had little to no effect on the rps.
Experiment info
The test has been done on a C3.4x AWS EC2 compute node (AMI image) with 16 vCPUs, with General SSD and 30GB RAM. During the test, CPU utilization peaked at 60% max (depends on the hatch rate - which controls the concurrent processes spawned), on an average staying under 30%.
Locust.io
setup: uses pyzmq, and setup with each vCPU core as a slave. Single POST request setup with request body ~ 20 bytes, and response body ~ 25 bytes. Request failure rate: < 1%, with mean response time being 6ms.
variables: Time between consecutive requests set to 450ms (min:100ms and max: 1000ms), hatch rate at a comfy 30 per sec, and RPS measured by varying #_users.
The RPS follows the equation as predicted for upto 1000 users. Increasing #_users after that has diminishing returns with a cap reached at roughly 1200 users. #_users here isn't the independent variable, changing the wait time affects the RPS as well. However, changing the experiment setup to 32 cores instance (c3.8x instance) or 56 cores (in a distributed setup) doesn't affect the RPS at all.
So really, what is the way to control the RPS? Is there something obvious I am missing here?
(one of the Locust authors here)
First, why do you want to control the RPS? One of the core ideas behind Locust is to describe user behavior and let that generate load (requests in your case). The question Locust is designed to answer is: How many concurrent users can my application support?
I know it is tempting to go after a certain RPS number and sometimes I "cheat" as well by striving for an arbitrary RPS number.
But to answer your question, are you sure your Locusts doesn't end up in a dead lock? As in, they complete a certain number of requests and then become idle because they have no other task to perform? Hard to tell what's happening without seeing the test code.
Distributed mode is recommended for larger production setups and most real-world load tests I've run have been on multiple but smaller instances. But it shouldn't matter if you are not maxing out the CPU. Are you sure you are not saturating a single CPU core? Not sure what OS you are running but if Linux, what is your load value?
While there is no direct way of controlling rps, you can try constant_pacing and constant_throughput option in wait_time
From docs
https://docs.locust.io/en/stable/api.html#locust.wait_time.constant_throughput
In the following example the task will always be executed once every 1 seconds, no matter the task execution time:
class MyUser(User):
wait_time = constant_throughput(1)
constant_pacing is inverse of this.
So if you run with 100 concurrent users, test will run at 100rps (assuming each request takes less than 1 second in first place
I have a time-critical application which processes a sequence of images coming from camera. It is written in C++ and it uses Qt, OpenCV and boost libraries. It is going to run on a dedicated PC.
Currently, the gui functions in main thread and i open a new thread for image processing. I didn't bother to divide the process section into threads because i think OpenCV is already doing that. However, i am having trouble maintaining the maximum tolerable delay.
My question is, how can i learn if my application using all the cores in the maximum level ?
When i look at the performance monitor, the pattern i see is really strange. The CPU usage is likely %35-40, all the cores are working but not at a full throttle.
Am i doing something wrong ?
You are not doing anything wrong, however you could change your code to take full use of the cpu cores by:
1 - setting the core affinity so that the thread does not change from one core to another, this could improve the cache usage (L1 and maybe L2)
2 - setting the scheduling of threads to FIFO so it does not get context-switched before finishing its processing
3 - run that thread on a higher priority process (this would require root privilege for the process)
Cheers