I'm seeing decreasing performance with each subsequent epoch running Tensorflow-GPU with Keras on Windows.
I'm training a 16-layer CNN and load my data using tf.data. The first epoch performs as well as expected: ~1hr training time. The CPU and GPU CUDA load is at 85-90%. Temperatures are reasonable (CPU between 65-70C and GPU at 70C) -- no thermal throttling occurs.
But by the second epoch, GPU CUDA load inexplicably drops to 33-50%. CPU load appears reduced as well, to around 65-75%. I don't see any big change in disk I/O speeds (the data is being loaded from an NVMe SSD).
I don't think this is a hardware issue. I'm using an RTX 3090 and a 4-core i5-6500 CPU and as mentioned, the performance on the first epoch is quite good without enough thermal headroom to continue. My tf.data pipeline looks like this:
# Construct tf.data.Dataset
paths_ds = tf.data.Dataset.from_tensor_slices(image_paths)
if shuffle:
paths_ds = paths_ds.shuffle(buffer_size=len(image_paths), reshuffle_each_iteration=True) # reshuffle each epoch
dataset = paths_ds.map(
lambda path: (self.load_image(path, scale_range, crop_size, augment), self.get_onehot_label(path, tf_class_index_by_image, self.num_classes)),
num_parallel_calls=tf.data.experimental.AUTOTUNE
)
dataset = dataset.batch(batch_size)
dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
A map and a prefetch, both using AUTOTUNE. I suspect maybe AUTOTUNE is the problem but I'm not sure how to debug this and determine how the number of threads and prefetch buffer size might be evolving over time. There is no documentation on how AUTOTUNE works.
EDIT: I did some profiling and have discovered that from epoch 2 onwards, Iterator::MapAndBatch is consuming significant time (it hard to even see in Epoch 1 because the data it collects is always ready). The cause appears to be some sort of perplexing slowness in my "load_image" function. Each operation it performs (read file, decode JPEG, convert to float, resize) is followed by a pause in Epoch 2 onwards and this becomes slower over time.
For example, Epoch 1 with everything functioning correctly:
Note the read is immediately followed by a decode, etc.
Now in epoch 2:
What could be causing this? Attempting to elevate the priority of the process using Process Explorer has no effect.
Related
I am trying out one huggingface sample with SWAG dataset
https://github.com/huggingface/transformers/tree/master/examples/pytorch/multiple-choice
I would like to use Intel Extension for Pytorch in my code to increase the performance.
Here I am using the one without training (run_swag_no_trainer)
In the run_swag_no_trainer.py , I made some changes to use ipex .
#Code before changing is given below:
device = accelerator.device
model.to(device)
#After adding ipex:
import intel_pytorch_extension as ipex
device = ipex.DEVICE
model.to(device)
While running the below command, its taking too much time.
export DATASET_NAME=swag
accelerate launch run_swag_no_trainer.py \
--model_name_or_path bert-base-cased \
--dataset_name $DATASET_NAME \
--max_seq_length 128 \
--per_device_train_batch_size 32 \
--learning_rate 2e-5 \
--num_train_epochs 3 \
--output_dir /tmp/$DATASET_NAME/
Is there any other method to test the same on intel ipex?
First you have to understand, which factors actually increases the running time. Following are these factors:
The large input size.
The data structure; shifted mean, and unnormalized.
The large network depth, and/or width.
Large number of epochs.
The batch size not compatible with physical available memory.
Very small or high learning rate.
For fast running, make sure to work on the above factors, like:
Reduce the input size to the appropriate dimensions that assures no loss in important features.
Always preprocess the input to make it zero mean, and normalized it by dividing it by std. deviation or difference in max, min values.
Keep the network depth and width that is not to high or low. Or always use the standard architecture that are theoretically proven.
Always make sure of the epochs. If you are not able to make any further improvements in your error or accuracy beyond a defined threshold, then there is no need to take more epochs.
The batch size should be decided based on the available memory, and number of CPUs/GPUs. If the batch cannot be loaded fully in memory, then this will lead to slow processing due to lots of paging between memory and the filesystem.
Appropriate learning rate should be determine by trying multiple, and using that which gives the best reduction in error w.r.t. number of epochs.
I have an Intel(R) Core(TM) i7-4720HQ CPU # 2.60GHz (Haswell) processor. In a relatively idle situation, I ran the following Perf commands and their outputs are shown, below. The counters are offcore_response.all_data_rd.l3_miss.any_response and mem_load_uops_retired.l3_miss:
sudo perf stat -a -e offcore_response.all_data_rd.l3_miss.any_response,mem_load_uops_retired.l3_miss sleep 10
Performance counter stats for 'system wide':
3,713,037 offcore_response.all_data_rd.l3_miss.any_response
2,909,573 mem_load_uops_retired.l3_miss
10.016644133 seconds time elapsed
These two values seem consistent, as the latter excludes prefetch requests and those not targeted at DRAM. But they do not match the read counter in the IMC. This counter is called UNC_IMC_DRAM_DATA_READS and documented here. I read the counter reread it 1 second later. The difference was around 30,000,000 (EDITED). If multiplied by 10 (to estimate for 10 seconds) the resulting value will be around 300 million (EDITED), which is 100 times the value of the above-mentioned performance counters (EDITED). It is nowhere near 3 million! What am I missing?
P.S.: The difference is much smaller (but still large), when the system has more load.
The question is also asked, here:
https://community.intel.com/t5/Software-Tuning-Performance/Performance-Counters-and-IMC-Counter-Not-Matching/m-p/1288832
UPDATE:
Please note that PCM output matches my IMC counter reads.
This is the relevant PCM output:
The values for columns READ, WRITE and IO are calculated based on UNC_IMC_DRAM_DATA_READS, UNC_IMC_DRAM_DATA_WRITES and UNC_IMC_DRAM_IO_REQUESTS, respectively. It seems that requests classified as IO will be either READ or WRITE. In other words, during the depicted one second interval, almost (because of the inaccuracy reported in the above-mentioned doc) 2.01GB of the 2.42GB READ and WRITE requests belong to IO. Based on this explanation, the above three columns seem consistent with each other.
The problem is that there still exists a LARGE gap between the IMC and PMC values!
The situation is the same when I boot in runlevel 1. The processes on the scheduler are one of swapper, kworker and migration. Disk IO is almost 85KB/s. I'm wondering what leads to such a (relatively) huge amount of IO. Is it possible to detect that (e.g., using a counter or a tool)?
UPDATE 2:
I think that there is something wrong with the IO column. It is always something in the range [1.99,2.01], regardless of the amount of load in the system!
UPDATE 3:
In runlevel 1, the average number of occurrences of the uops_retired.all event in a 1-second interval is 15,000,000. During the same period, the number of read requests recorded by the associated IMC counter is around 30,000,000. In other words, assuming that all memory accesses are directly caused by cpu instructions, for each retired micro-operation, there exists two memory accesses. This seems impossible specially concerning the fact that there exist multiple levels of caches. Therefore, in the idle scenario, perhaps, the read accesses are caused by IO.
Actually, it was mostly caused by the GPU device. This was the reason for exclusion from performance counters. Here is the relevant output for a sample execution of PCM on a relatively idle system with resolution 3840x2160 and refresh rate 60 using xrandr:
And this is for the situation with resolution 800x600 and the same refresh rate (i.e., 60):
As can be seen, changing screen resolution reduced read and IO traffic considerably (more than 100x!).
I am running a Python program that calls H2O for deep learning (training and testing). The program runs in a loop of 20 iterations and in each loop calls H2ODeepLearningEstimator() 4 times and associated predict() and model_performance(). I am doing h2o.remove_all() and cleaning up all data-related Python objects after each iteration.
Data size: training set 80,000 with 122 features (all float) with 20% for validation (10-fold CV). test set 20,000. Doing binary classification.
Machine 1: Windows 7, 4 core, Xeon, each core 3.5GHz, Memory 32 GB
Takes about 24 hours to complete
Machine 2: CentOS 7, 20 core, Xeon, each core 2.0GHz, Memory 128 GB
Takes about 17 hours to complete
I am using h2o.init(nthreads=-1, max_mem_size = 96)
So, the speed-up is not that much.
My questions:
1) Is the speed-up typical?
2) What can I do to achieve substantial speed-up?
2.1) Will adding more cores help?
2.2) Are there any H2O configuration or tips that I am missing?
Thanks very much.
- Mohammad,
Graduate student
If the training time is the main effort, and you have enough memory, then the speed up will be proportional to cores times core-speed. So, you might have expected a 40/14 = 2.85 speed-up (i.e. your 24hrs coming down to the 8-10 hour range).
There is a typo in your h2o.init(): 96 should be "96g". However, I think that was a typo when writing the question, as h2o.init() would return an error message. (And H2O would fail to start if you'd tried "96", with the quotes but without the "g".)
You didn't show your h2o.deeplearning() command, but I am guessing you are using early stopping. And that can be unpredictable. So, what might have happened is that your first 24hr run did, say, 1000 epochs, but your second 17hr run did 2000 epochs. (1000 vs. 2000 would be quite an extreme difference, though.)
It might be that you are spending too much time scoring. If you've not touched the defaults, this is unlikely. But you could experiment with train_samples_per_iteration (e.g. set it to 10 times the number of your training rows).
What can I do to achieve substantial speed-up?
Stop using cross-validation. That might be a bit controversial, but personally I think 80,000 training rows is going to be enough to do an 80%/10%/10% split into train/valid/test. That will be 5-10 times quicker.
If it is for a paper, and you want to show more confidence in the results, once you have your final model, and you've checked that test score is close to valid score, then rebuild it a couple of times using a different seed for the 80/10/10 split, and confirm you end up with the same metrics. (*)
*: By the way, take a look at the score for each of the 10 cv models you've already made; if they are fairly close to each other, then this approach should work well. If they are all over the place, you might have to re-consider the train/valid/test splits - or just think about what it is in your data that might be causing that sensitivity.
I've been working on an Arduino (ATMega328p) prototype that has to log data during certain events. An LSM6DS33 sensor is used to generate 6 values (2 bytes each) at a sample rate of 104 Hz. This data needs to be logged for a period of 500-20000ms.
In my code, I generate an interrupt every 1/104 sec using Timer1. When this interrupt occurs, data is read from the sensor, calibrated and then written to an SD card. Normally, this is not an issue. Reading the data from the sensor takes ~3350us, calibrating ~5us and writing ~550us. This means a total cycle takes ~4000us, whereas 9615us is available.
In order to save power, I wish to lower the voltage to 3.3V. According to the atmel datasheet, this also means that the clock frequency should be lowered to 8MHz. Assuming everything will go twice as slow, a measurement cycle would still be possible because ~8000us < 9615us.
After some testing (still 5V#16MHz), however, it occured to me that every now and then, a write cycle would take ~1880us instead of ~550us. I am using the library SdFat to write and test SD cards (RawWrite example). The following results came in when I tested the card:
Start raw write of 100000 KB
Target rate: 100 KB/sec
Target time: 100 seconds
Min block write time: 1244 micros
Max block write time: 12324 micros
Avg block write time: 1247 micros
As seen, the average time to write is fairly consistent, but sometimes a peak duration of 10x average occurs! According to the writer of the library, this is because the SD card needs some erase cycles in between x amount of write cycles. This causes a write delay (src:post#18). This delay, however, pushes the time required for a cycle out of the available 9615us bracket, because the total measure cycle would be 10672us.
The data I am trying to write, is first put into a string using sprintf:
char buf[20] = "";
sprintf(buf,"%li\t%li\t%li\t%li\t%li\t%li",rawData[0],rawData[1],rawData[2],rawData[3],rawData[4],rawData[5]);
myLog.println(buf);
This writes the data to a txt file. But at my speed rate, only 21*104=2184 B/s would suffice. Lowering the speed of the RawWrite example to 6 KB/s, causes the SD card to write without getting an extended write delay. Yet my code still has them, even though less data is written.
My question is: how do I prevent this delay from occurring (if possible)? And if not possible, how can I work around it? It would help if I understood why exactly the delay occurs, because the interval is not always the same (every 10-15 writes).
Some additional info:
The sketch currently uses 69% of RAM (2kB) with variables. Creating two 512 byte buffers - like suggested in the same forum - is not possible for me.
Initially, I used two strings. Merging them into one, didn't affect the write speed with any significance.
I don't know how to work around the delay, but I experience a more stable and faster writing time, if I wrote to a binary file instead of a ".csv" or .txt" file.
The following link provide a fine script to write data as a binary struct to the SD card. (There are some small typo in his example, it is easily fixed)
https://hackingmajenkoblog.wordpress.com/2016/03/25/fast-efficient-data-storage-on-an-arduino/
This will not help you with the time variation, but it might minimize the writing time, and thus negleting the time issue.
How do I find the optimal chunk size for multiprocessing.Pool instances?
I used this before to create a generator of n sudoku objects:
processes = multiprocessing.cpu_count()
worker_pool = multiprocessing.Pool(processes)
sudokus = worker_pool.imap_unordered(create_sudoku, range(n), n // processes + 1)
To measure the time, I use time.time() before the snippet above, then I initialize the pool as described, then I convert the generator into a list (list(sudokus)) to trigger generating the items (only for time measurement, I know this is nonsense in the final program), then I take the time using time.time() again and output the difference.
I observed that the chunk size of n // processes + 1 results in times of around 0.425 ms per object. But I also observed that the CPU is only fully loaded the first half of the process, in the end the usage goes down to 25% (on an i3 with 2 cores and hyper-threading).
If I use a smaller chunk size of int(l // (processes**2) + 1) instead, I get times of around 0.355 ms instead and the CPU load is much better distributed. It just has some small spikes down to ca. 75%, but stays high for much longer part of the process time before it goes down to 25%.
Is there an even better formula to calculate the chunk size or a otherwise better method to use the CPU most effective? Please help me to improve this multiprocessing pool's effectiveness.
This answer provides a high level overview.
Going into detais, each worker is sent a chunk of chunksize tasks at a time for processing. Every time a worker completes that chunk, it needs to ask for more input via some type of inter-process communication (IPC), such as queue.Queue. Each IPC request requires a system call; due to the context switch it costs anywhere in the range of 1-10 μs, let's say 10 μs. Due to shared caching, a context switch may hurt (to a limited extent) all cores. So extremely pessimistically let's estimate the maximum possible cost of an IPC request at 100 μs.
You want the IPC overhead to be immaterial, let's say <1%. You can ensure that by making chunk processing time >10 ms if my numbers are right. So if each task takes say 1 μs to process, you'd want chunksize of at least 10000.
The main reason not to make chunksize arbitrarily large is that at the very end of the execution, one of the workers might still be running while everyone else has finished -- obviously unnecessarily increasing time to completion. I suppose in most cases a delay of 10 ms is a not a big deal, so my recommendation of targeting 10 ms chunk processing time seems safe.
Another reason a large chunksize might cause problems is that preparing the input may take time, wasting workers capacity in the meantime. Presumably input preparation is faster than processing (otherwise it should be parallelized as well, using something like RxPY). So again targeting the processing time of ~10 ms seems safe (assuming you don't mind startup delay of under 10 ms).
Note: the context switches happen every ~1-20 ms or so for non-real-time processes on modern Linux/Windows - unless of course the process makes a system call earlier. So the overhead of context switches is no more than ~1% without system calls. Whatever overhead you're creating due to IPC is in addition to that.
Nothing will replace the actual time measurements. I wouldn't bother with a formula and try a constant such as 1, 10, 100, 1000, 10000 instead and see what works best in your case.