Some documents may need to be updated

[DylanWangWQF]

Hi @jlwatson I tried to run piranha locally on 3 GPUs and build it in the docker container.

1. In Makefile, CUDA_VERSION=11.5 while the installed version is specified as11.6 in Dockerfile. I guess here needs the synchronization in piranha (I just changed the version in Makefileto avoid make error).

2. After solving several configuration issues, I can run piranha locally now. I guess some commands in the Dockfile should be updated (e.g., download the dataset for training, etc..). Here is the running log:

   root@a5cb7eceb1f6:/piranha# ./localhost_runner.sh 
   run unit tests? false
   config network: "files/models/secureml-norelu.json"
   network filename: files/models/secureml-norelu.json
   ----------------------------------------------
   (1) FC Layer          784 x 128
             512        (Batch Size)
   ----------------------------------------------
   (2) ReLU Layer        512 x 128
   ----------------------------------------------
   (3) FC Layer          128 x 128
             512        (Batch Size)
   ----------------------------------------------
   (4) ReLU Layer        512 x 128
   ----------------------------------------------
   (5) FC Layer          128 x 10
             512        (Batch Size)
   Error opening training data file at files/MNIST/train_data
   Error opening training label file at files/MNIST/train_labels
   Error opening test data file at files/MNIST/test_data
   Error opening test label file at files/MNIST/test_label
   TRAINING, EPOCHS = 10 ITERATIONS = 117
   epoch,0
   total time (s),63.125727
   total tx comm (MB),2339.085938
   total rx comm (MB),2339.085938
   train accuracy,0.000000
   epoch,1
   total time (s),126.424196
   total tx comm (MB),4678.171875
   total rx comm (MB),4678.171875
   train accuracy,0.000000

Here are some logs in my machine if someone wants to check:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.47.03    Driver Version: 510.47.03    CUDA Version: 11.6     |
+-------------------------------+----------------------+----------------------+

|   1  Tesla V100-DGXS...  On   | 00000000:08:00.0 Off |                    0 |
| N/A   38C    P0    53W / 300W |    371MiB / 32768MiB |     19%      Default |
+-------------------------------+----------------------+----------------------+
|   2  Tesla V100-DGXS...  On   | 00000000:0E:00.0 Off |                    0 |
| N/A   38C    P0    52W / 300W |    347MiB / 32768MiB |     18%      Default |
+-------------------------------+----------------------+----------------------+
|   3  Tesla V100-DGXS...  On   | 00000000:0F:00.0 Off |                    0 |
| N/A   39C    P0    54W / 300W |    339MiB / 32768MiB |     19%      Default |
+-------------------------------+----------------------+----------------------+

| Processes:                                                                                                             |
|  GPU   GI   CI        PID   Type   Process name       GPU Memory Usage|
|=================================================|
|    1   N/A  N/A      9249      C   ./piranha                         335MiB |
|    2   N/A  N/A      9250      C   ./piranha                         335MiB |
|    3   N/A  N/A      9252      C   ./piranha                         335MiB |
+-----------------------------------------------------------------------------+

[DylanWangWQF]

BTW, if we run piranha locally, we may need to add extra functions to simulate LAN and WAN environments. For me, I want to build my experiment based on piranha, so I need to collect the result under both LAN and WAN.

[jlwatson]

Hey @DylanWangWQF, coming back from out of town so apologies for the delay. I used Docker during some of my evaluation a while back, but haven't updated the Dockerfile for the current setup. Taking a look, since the Dockerfile copies your local files/ and scripts/ directories into the resulting image, you need to download the right dataset (e.g. MNIST) locally first.

That missing dataset might be why your local run above is not actually training.

[jlwatson]

As for LAN/WAN latency simulation, I wonder if it's a good idea to integrate that into Piranha or add that as a separate software layer somewhere else. In any case, I think it'll be hard to get the same performance running all parties locally because they're all contending for the PCIe bus to talk to their GPUs.

[DylanWangWQF]

Hey @DylanWangWQF, coming back from out of town so apologies for the delay. I used Docker during some of my evaluation a while back, but haven't updated the Dockerfile for the current setup. Taking a look, since the Dockerfile copies your local files/ and scripts/ directories into the resulting image, you need to download the right dataset (e.g. MNIST) locally first.

That missing dataset might be why your local run above is not actually training.

Yep! I have already tested it successfully.

[DylanWangWQF]

As for LAN/WAN latency simulation, I wonder if it's a good idea to integrate that into Piranha or add that as a separate software layer somewhere else. In any case, I think it'll be hard to get the same performance running all parties locally because they're all contending for the PCIe bus to talk to their GPUs.

Yes, if I run locally on a single machine, maybe I have to test how much it affects the overall performance of my exp. Thanks!

BTW, does piranha implement the functionality of the equality test in 3PC?