Colab question - Page 25

Uncwilly · 2020-11-21, 21:28

Quote:

Originally Posted by mrk74

Right a GPU. I'm going on about 10 days. Would it make a difference to switch the setting to TPU since we aren't getting a GPU anyway?

Do you have code that can take advantage of a TPU?

Ensigm · 2020-11-22, 13:01

Quote:

Originally Posted by mrk74

Right a GPU. I'm going on about 10 days. Would it make a difference to switch the setting to TPU since we aren't getting a GPU anyway?

I believe it has something to do with the time in a day. 22:00-01:00 GMT might give you a better chance to catch a GPU, especially a T4.

mrk74 · 2020-11-22, 15:55

Quote:

Originally Posted by Uncwilly

Do you have code that can take advantage of a TPU?

I don't know anything about coding at all so no.

FINALLY I catch a GPU! And a T4 at that. Edit: It only lasted 45 mins.

chalsall · 2020-11-24, 19:04

Hey everyone. Super busy at the moment; behind on an impossible schedule.

But... It definitely seems like the "honeymoon" is over with Colab.

I had a theory that each user was being treated differently by the "black box" that is Google. And the recent personal empirical observed by myself combined with other Colab users' observations support this.

My thirteen (13#) Colab accounts have been getting no more than an hour of GPU per day, usually only about 30 minutes, for the last week or so. And, for the last few days, even the CPU instances are now only surviving an hour or so.

Super weirdly, one GPU72 Colab Notebook user appears to continue to get four (4#) concurrent GPU sessions quite regularly.

While this is disappointing for our long-term resource management planning, we have absolutely no reason to complain.

It was fun while it lasted, we all learnt a huge amount about instance coding and IPC, and lastly, a huge amount of work was done!

Thanks, Google, for the playground and all the Factors. Sincerely.

P.S. I'm going to keep spinning up Colab instances as part of my daily workflow. I'll let everyone know if availability becomes more useful to us over time.

P.P.S. The anecdotal empirical suggests that those *not* running the GPU72 Colab Notebook might be having greater access. Excellent.

moebius · 2020-11-24, 19:33

I want to briefly explain my strategy with which I have from June 2020 until now to get to 95% a 10-12 hour gpu runtime.

1. Use of only 3 free google accounts

2. Exclusive use of only one notebook at the same time (log out at all other google accounts)

3. I only run gpuowl, no mprime, mfaktc or anything else and no P-1 with gpuowl

This leads to the following.

A. Always a GPU runtime available (mostly P100. T4, P4 always) with at least 1 of 3 accounts

B. Partial prevention of the download of files from Google Colab (Google Drive) to the PC

Bypassing B.

Release the file for unrestricted viewing / downloading for all.

May the google gods have mercy on me.

petrw1 · 2020-11-24, 19:38

=====> Is me.

Quote:

Originally Posted by moebius

I want to briefly explain my strategy with which I have from June 2020 until now to get to 95% a 10-12 hour gpu runtime.

1. Use of only 3 google accounts
=====> I have 2

2. Exclusive use of only one notebook at the same time (log out at all other google accounts)
=====> Both at same time

3. I only run gpuowl, no mprime, maktc or anything else
=====> Only GpuOwl latetly

This leads to the following.

A. Always a GPU runtime available (mostly P100. T4, P4 always) with at least 1 of 3 accounts
=====> 95% of the time it gives me a T4; eventually (5-25 tries) I get 2 P100's.

B. Partial prevention of the download of files from Google Colab to the PC

Bypassing B.

Release the file for unrestricted viewing / downloading for all.

May the google gods have mercy on me.

moebius · 2020-11-26, 23:52

Quote:

Originally Posted by petrw1

=====> Is me.
2. Exclusive use of only one notebook at the same time (log out at all other google accounts)
=====> Both at same time

Do you let the two accounts and notebooks run time-shifted or really at the same time? Otherwise you would at some point no longer have to get a runtime for both accounts.

petrw1 · 2020-11-27, 00:02

Quote:

Originally Posted by moebius

Do you let the two accounts and notebooks run time-shifted or really at the same time? Otherwise you would at some point no longer have to get a runtime for both accounts.

Usually both at the same time.
Occasionally I can't get two P100s in one sitting; I try again an hour or two later; and this may take 3 or 4 tries which of course staggers them a bit.
They are signed on with different emails so I'm hoping Colab thinks they are different people.

petrw1 · 2020-11-27, 22:03

For the last week or so I have been saving the UUID of every GPU Colab offers me.
In that time period of 125 T4 GPUs offered only two have the same UUID; and these were offered on different days; not back to back.

Aramis Wyler · 2020-11-28, 16:09

I noticed today that after being especially persistant about re-starting the sessions when they die, it started giving me a different error - that there simply were no gpus available. Not that I'd hit my quots, or couldn't connect to a gpu, but that there were none.

I pulled up the help on this and got a message about google policy strictly forbidding cryptocurrency mining, and so if were were doing that our gpu availability would be reduced or revoked altogether.

So that's what I think this is - it thinks we're doing cryptocurrency mining and removing gpu time accordingly.

EdH · 2020-11-28, 16:37

I tried one of my Colab "How I . . ." sessions with GPU and got a P100, but I haven't done any GPU runs in a few months. I never do try to get anything specific and I'm not too frequently running Colab, although recently I've had as many as four CPU sessions open on one account for all day (probably around 12 hours) a couple times.

2020-11-24, 19:04	#268
chalsall If I May "Chris Halsall" Sep 2002 Barbados 24C2₁₆ Posts	So long, and thanks for all the factors Colab... Hey everyone. Super busy at the moment; behind on an impossible schedule. But... It definitely seems like the "honeymoon" is over with Colab. I had a theory that each user was being treated differently by the "black box" that is Google. And the recent personal empirical observed by myself combined with other Colab users' observations support this. My thirteen (13#) Colab accounts have been getting no more than an hour of GPU per day, usually only about 30 minutes, for the last week or so. And, for the last few days, even the CPU instances are now only surviving an hour or so. Super weirdly, one GPU72 Colab Notebook user appears to continue to get four (4#) concurrent GPU sessions quite regularly. While this is disappointing for our long-term resource management planning, we have absolutely no reason to complain. It was fun while it lasted, we all learnt a huge amount about instance coding and IPC, and lastly, a huge amount of work was done! Thanks, Google, for the playground and all the Factors. Sincerely. P.S. I'm going to keep spinning up Colab instances as part of my daily workflow. I'll let everyone know if availability becomes more useful to us over time. P.P.S. The anecdotal empirical suggests that those not running the GPU72 Colab Notebook might be having greater access. Excellent.

2020-11-24, 19:33	#269
moebius Jul 2009 Germany 547 Posts	I want to briefly explain my strategy with which I have from June 2020 until now to get to 95% a 10-12 hour gpu runtime. 1. Use of only 3 free google accounts 2. Exclusive use of only one notebook at the same time (log out at all other google accounts) 3. I only run gpuowl, no mprime, mfaktc or anything else and no P-1 with gpuowl This leads to the following. A. Always a GPU runtime available (mostly P100. T4, P4 always) with at least 1 of 3 accounts B. Partial prevention of the download of files from Google Colab (Google Drive) to the PC Bypassing B. Release the file for unrestricted viewing / downloading for all. May the google gods have mercy on me. Last fiddled with by moebius on 2020-11-24 at 20:06

2020-11-27, 22:03	#273
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 13·349 Posts	Do we keep getting the same GPU? .... almost never. For the last week or so I have been saving the UUID of every GPU Colab offers me. In that time period of 125 T4 GPUs offered only two have the same UUID; and these were offered on different days; not back to back.

2020-11-28, 16:09	#274
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 3×137 Posts	I noticed today that after being especially persistant about re-starting the sessions when they die, it started giving me a different error - that there simply were no gpus available. Not that I'd hit my quots, or couldn't connect to a gpu, but that there were none. I pulled up the help on this and got a message about google policy strictly forbidding cryptocurrency mining, and so if were were doing that our gpu availability would be reduced or revoked altogether. So that's what I think this is - it thinks we're doing cryptocurrency mining and removing gpu time accordingly.

2020-11-28, 16:37	#275
EdH "Ed Hall" Dec 2009 Adirondack Mtns DDA₁₆ Posts	I tried one of my Colab "How I . . ." sessions with GPU and got a P100, but I haven't done any GPU runs in a few months. I never do try to get anything specific and I'm not too frequently running Colab, although recently I've had as many as four CPU sessions open on one account for all day (probably around 12 hours) a couple times.