How to reproduce BERT perf results in deepspeed blog

[LiweiPeng]

Hi Deepspeed team,

The BERT perf results from blog Microsoft DeepSpeed achieves the fastest BERT training time are very impressive. However I couldn't reproduce the perf results in Figure 1 in the blog.

For example, using the latest nvbert code, for bert large model max-seq-len=128, the max batch size I got is 136 with 194.57 examples/s. However, in figure1, nvbert's max batch size is about 82 with 215 examples/s. The nvbert perf is 10% better than what I got.

Can you share the detailed parameters (and/or code) for Figure 1 in the blog on how to reproduce the nvbert perf results and huggingface bert perf results? Thanks.

Thanks.
Liwei

[RezaYazdaniAminabadi]

Hi Liwei,

Thanks for trying out the deepspeed transformer kernel.
I assume you use a 32-GB NVIDIA V100 GPU for this experiment, am I right?
I wonder what total batch size you are using. I mean if it is with gradient accumulation of 1 or higher. Also, is it a post-ln or pre-ln BERT layer that you are running the training with?
We use the pre-ln BERT structure and we set the gradient accumulation to 10 for all our experiments.
One last thing is that which dataset you are using for the training?

Thanks.
Reza

[LiweiPeng]

@RezaYazdaniAminabadi , Thanks for your quick response.

To answer your questions:

• I am using one 32-GB Nvidia V100 GPU.
• I am using the post-layernorm in nvidia's code because this is the only mode nvbert supports.
• I am using English dataset (wikipedia + bookcorpus).
• I tried 2 sets of test:
  test1: max batch=136, gradient accumulation=1, perf is 194.57 examples/s.
  test2: max batch=132, gradient accumulation=10, perf is 200.50 examples/s.

One difference from your test is that I am using post-ln not pre-ln. I am not sure whether this can explain the max-batch size difference and perf difference or not. I'll try pre-ln later.

Thanks.
Liwei

[RezaYazdaniAminabadi]

Thanks for the reply. If you want to try pre-ln model, you may use the example in deepspeed: https://github.com/microsoft/DeepSpeedExamples/blob/8610e5e3fcce5fb247e3b85ea2bed0f2296b5443/bing_bert/nvidia/modelingpreln.py
Regarding the performance difference, the pre-ln may have some performance difference than post-ln. Another difference is that we use deepspeed to run the nvidia modeling. You can try the deepspeed example:
https://github.com/microsoft/DeepSpeedExamples/tree/8610e5e3fcce5fb247e3b85ea2bed0f2296b5443/bing_bert
To run the pretraining, you can use ds_train_bert_bsz64k_seq128.sh script and change the batch size in deepspeed_bsz64k_lamb_config_seq128.json.

[LiweiPeng]

Thanks Reza. I'll try the deepspeed example way.

[JF-D]

@RezaYazdaniAminabadi I tried DeepSpeed bing_bert example, but I only got 200 samples/sec with batch size 64 and gradient accumulation of 10 for 128 max sequence length. I use a 32GB V100 GPU and dummy dataloader. I am quite confused about this result.

Besides, for distributed BERT Large training, I discovered that the communication time is quite long. I wonder how did you get that high speedup ration.

-- update
by enabling attention dropout checkpoint and normalize invertible in transformer kernel, I got 5% gain. But it's still far away from the deepspeed result.

[JF-D]

@LiweiPeng Hi, did you try the deepspeed example and get the same perf results in deepspeed blog?

[LiweiPeng]

@JF-D I haven't got a chance to try the deepspeed example yet. will update once I try it.

[RezaYazdaniAminabadi]

Hi @JF-D,
Thanks for trying out the deepspeed transformer kernel.
I just ran the DeepSpeedExample with setting the --deepspeed_transformer_kernel flag in bash ds_train_bert_bsz64k_seq128.sh script. Also, I changed the train_batch_size from 64K to 640 (considering gradient accumulation of 10) in deepspeed json config file. I am using the rest of the configuration as default. I can get much higher samples/second, around 257. I achieve this result without setting any other optimization flags.

You can further increase the performance by setting the flags you mentioned and increase the batch size.
Can you please share more information for your run? Like, what configuration you are using and also the script you run training with?
For the distributed training, can you please set the wall_clock_breakdown flag in deepspeed config to true? This way, we can look into more detail of your training system, like what are the forward, backward and all-reduce time.
After I get this information from you, I can help you to get the desired performance from deepspeed! :)

Thanks.
Reza

[RezaYazdaniAminabadi]

By the way, if you even set the gradient accumulation to 1, using 1 NVIDIA V100 GPU (32GB), you should get a high samples/sec.

[JF-D]

@RezaYazdaniAminabadi Thanks for your quick response.
I just follow your configuration except that I made some changes to the dataloader. But I used a dummy dataloader to avoid its effects.

I just ran the DeepSpeedExample with setting the --deepspeed_transformer_kernel flag in bash ds_train_bert_bsz64k_seq128.sh script. Also, I changed the train_batch_size from 64K to 640 (considering gradient accumulation of 10) in deepspeed json config file. I am using the rest of the configuration as default.

After setting the wall_clock_breakdown flag in deepspeed config to true, I get this log.

For other info, I use pytorch 1.5, cuda 10.1, cudnn 7.6.5, nccl 2.6.4.
And here is my GPU info.

[RezaYazdaniAminabadi]

Hi @JF-D,

Thanks for providing more information of your experiment.
Clearly, we have some hardware differences here. I am using a DGX-2. Also, I am performing the evaluation using Pytorch 1.2. You can find the GPU info as follows. In my experiment, I am seeing the power is on average around 340 Watt, and the GPU utilization is >98%.

The FWD/BWD time you've got is almost 11.5% higher than what I am seeing:

Also, I realize you made some changes at the data-loader. To speed up the data loading, we use an asynchronous worker on CPU which prefetches the data in advance. So, my suggestion is that please use a fresh version of our DeepSpeed example and please check in the new changes of deepspeed, and try the experiment again. Let's see if your numbers improve! :)
Thank.

By the way, you may want to turn on the --stochastic-mode in the training script to boost the performance of the transformer kernel.

Best regards,
Reza

[JF-D]

@RezaYazdaniAminabadi Thanks a lot. I have a few more questions and I look forward to receiving your answers.

1. It seems that the difference between our speed is caused by the different hardware (V100 SXM3 vs SXM2). I don't have dgx2 and I don't know much about their difference in training networks. Do you think that this difference is normal, and dgx2 can be 10% faster?
2. For your training, did you use NVMe? I didn't use NVMe, but it seems that using nvme can get some performance gain. I am not sure.
3. I discover that my gpu doesn't seem to be fully utilized (the GPU utilization sometimes < 50%). There are occasional fluctuations in the speed of the log and the time becomes longer. I think this is abnormal, there may be some problems with the machine, right?
   
4. I am quite interested in your allreduce time for large scale distributed training. The BERT Large model have ~330m parameters. Thus, for each iteration, the communication volume is ~660MB. Deepspeed can train BERT Large in 44mins with 1024 GPUs and that's an amazing result. Did you do other optimizations for allreduce? In my opinion, allreduce time is a big bottleneck in large-scale scenario.

Looking forward to your reply!

Thanks,
Jiangfei

[TonyTangYu]

Hi, @LiweiPeng

I am trying to run the examples on the dataset wikipedia and bookcorpus. However, the official deepspeed website says that the Downloading and pre-processing instructions of these datasets are coming soon. I notice that you conduct some experiments on these two datasets. Could you please offer me some pre-processing instructions of these two datasets for me to run the DeepSpeed/BERT Pre-training examples?

Thanks!
Tony

[tjruwase]

@TonyTangYu, can you try the nvidia datasets? Please see relevant information here and here.

[TonyTangYu]

@tjruwase ， thanks for your reply! I will give it a try!
Thanks again!