Google Diet Colab Notebook

[xx005fs]

It turns out the K80 is significantly faster than the T4 (yesterday didn't manage to get an actual T4 instance to benchmark the speed difference). Generally I can get around 4.5ms/it with the K80 while the T4 is about 6ms/it after it starts to throttle down. Here are the results

K80:

2019-09-08 16:32:02 90396263 OK 27360000 30.27%; 4559 us/sq; ETA 3d 07:49; 3ec593b85e44fb66 (check 1.12s)
2019-09-08 16:35:06 90396263 OK 27400000 30.31%; 4561 us/sq; ETA 3d 07:49; ddc1e8d47986dac7 (check 1.11s)
2019-09-08 16:38:09 90396263 OK 27440000 30.36%; 4561 us/sq; ETA 3d 07:45; d7ad382a0b7037d3 (check 1.11s)
2019-09-08 16:41:13 90396263 OK 27480000 30.40%; 4560 us/sq; ETA 3d 07:42; 30103cf3945858fc (check 1.10s)
2019-09-08 16:44:17 90396263 OK 27520000 30.44%; 4561 us/sq; ETA 3d 07:39; 496945fc83650272 (check 1.11s)

T4:

2019-09-08 17:04:58 90396473 OK 27625600 30.56%; 5578 us/sq; ETA 4d 01:15; eb92599fdef067db (check 1.29s)
2019-09-08 17:06:24 90396473 OK 27640000 30.58%; 5884 us/sq; ETA 4d 06:35; ec8ff248229a7a0d (check 1.41s)
2019-09-08 17:10:27 90396473 OK 27680000 30.62%; 6038 us/sq; ETA 4d 09:11; 62cd89f65e138ebc (check 1.38s)
2019-09-08 17:14:30 90396473 OK 27720000 30.66%; 6038 us/sq; ETA 4d 09:08; f9c1164a0935332f (check 1.38s)

[hansl]

Originally Posted by xx005fs

Similarly, I am using... Colab to run gpuowl

Just curious why gpuowl and not CUDALucas? I haven't tried either yet so I hardly know anything about them, but I guess I would assume that CUDA would generally beat OpenCL on nvidia GPUs?

[xx005fs]

Originally Posted by hansl

Just curious why gpuowl and not CUDALucas? I haven't tried either yet so I hardly know anything about them, but I guess I would assume that CUDA would generally beat OpenCL on nvidia GPUs?

I personally think the fact that PRP has a reliable error check algorithm makes it superior to LL. Secondly, gpuowl seems to run faster on Nvidia cards that are bandwidth starved (for example my Titan V is significantly faster on gpuowl than CUDALucas, down from 1.12ms/it to 0.83ms/it with the switch. The Tesla K80 is no exception since it has a 1:3 FP32:FP64 rate but a relatively low memory bandwidth, the same should apply.), and obviously I would want maximum throughput. I would try CUDALucas sometimes if it's faster than gpuowl on the K80.

[chalsall]

Originally Posted by Corbeau

Be careful using more than one Google account. Apparently people on the LCZero project were banned from using CoLab because they did that.

I would really like to get a handle on what Google considers fair-use for this kind of thing. I'll reach out to them to try to get some clarity.

But... Doing some research on this I came across another Google offering: Kaggle. I registered (using my Google account), and after authenticating by way of a SMS message I was given a very similar Notebook environment.

Pasted in my beta Notebook and clicked run, and it happily gave me a Tesla P100-PCIE-16GB producing ~1,200 GHzD/D of TF'ing -- no changes to my code. Clicked on "Commit and Run" and I was allowed to launch two more batch runs, each with another P100!

The interactive session is limited to nine (9) hours, while the batch runs are limited to six (6). They also provide 5 GB of persistent storage per Notebook.

Like, wow!

[xx005fs]

Originally Posted by chalsall

I would really like to get a handle on what Google considers fair-use for this kind of thing. I'll reach out to them to try to get some clarity.

But... Doing some research on this I came across another Google offering: Kaggle. I registered (using my Google account), and after authenticating by way of a SMS message I was given a very similar Notebook environment.

Pasted in my beta Notebook and clicked run, and it happily gave me a Tesla P100-PCIE-16GB producing ~1,200 GHzD/D of TF'ing -- no changes to my code. Clicked on "Commit and Run" and I was allowed to launch two more batch runs, each with another P100!

The interactive session is limited to nine (9) hours, while the batch runs are limited to six (6). They also provide 5 GB of persistent storage per Notebook.

Like, wow!

That's amazing, the P100 is an amazing GPU for PRP/LL. I would rather run those workloads instead of TF since that should be left to the Turing GPUs. Any tips on how to set it up since I can't seem to run gpuowl from uploading files.

[Dylan14]

Originally Posted by Dylan14

A true acid test would be to get BOINC set up on the notebook. It would prove handy for quite a lot of projects (thinking NFS@Home with the Xeon, or the PPS sieve on the GPU for PrimeGrid).

Well, butter my biscuit, I have created code for this. This took a bit of tinkering, but here it is:

import os.path
#Use apt-get to get boinc
!apt-get install boinc boinc-client
#cp boinc, boinccmd to working directory
!cp /usr/bin/boinc /content
!cp /usr/bin/boinccmd /content
#create a slots directory if it doesn't exist(otherwise boinc doesn't work)
if not os.path.exists('/content/slots'):
  !mkdir slots
#launch the client
#attach to projects as desired (here I used NFS@home)
if not os.path.exists('/content/slots/0'):
  !boinc --attach_project https://escatter11.fullerton.edu/nfs/ (your account key here)
else:
  !boinc

replace (your account key here) with your actual account key. Note you'll have to register for the project first. Also, the processor will default to the blank location: make sure you have the preferences set correctly!

[chalsall]

Originally Posted by Dylan14

Well, butter my biscuit, I have created code for this. This took a bit of tinkering, but here it is:

Absolutely mind-blowingly awesome!!!

I have reached out to both Colaboratory and Kaggle, pointing to this thread and asking them if what we're doing here is considered OK by them, and if they have any comments or feedback.

I consider this to be much like running compute on an employer's or client's kit -- best to get explicit permission.

[LaurV]

grrr... now because of you, they will nerf it...

[GP2]

FAQ says:

What is Colaboratory?
Colaboratory is a research tool for machine learning education and research.

Why are hardware resources such as T4 GPUs not available to me?
The best available hardware is prioritized for users who use Colaboratory interactively rather than for long-running computations. Users who use Colaboratory for long-running computations may be temporarily restricted in the type of hardware made available to them, and/or the duration that the hardware can be used for. We encourage users with high computational needs to use Colaboratory’s UI with a local runtime.
Please note that using Colaboratory for cryptocurrency mining is disallowed entirely, and may result in being banned from using Colab altogether.

So it sounds like they somewhat tolerate compute-intensive usage even though the purpose of the thing is machine learning.

[chalsall]

Originally Posted by GP2

So it sounds like they somewhat tolerate compute-intensive usage even though the purpose of the thing is machine learning.

I haven't heard back from either of them, and I continue to be allowed to compute, so...

I could really use a couple of more beta testers. If you have a GPU72 account and a Google account, and are willing to help out, please PM me. It's pretty simple.

As always, you get all the credit for the work done on your behalf.

[De Wandelaar]

For the Kaggle user's, see https://www.kaggle.com/general/108481.
From now on, GPU usage is limited to 30 hours per week.