mersenneforum.org Colab question
 Register FAQ Search Today's Posts Mark Forums Read

2020-11-07, 02:03   #243
RichD

Sep 2008
Kansas

7·463 Posts

Time to put an end to this nonsense. It was a nice ride, nearly 20 fold in stats. I get more from a free account than a paid. There is no free lunch. I'll continue my local GTX 560 Ti farm for a while.
Attached Files
 GPU.pdf (52.1 KB, 38 views)

Last fiddled with by RichD on 2020-11-07 at 02:06 Reason: added attachment

2020-11-07, 21:46   #244
Mark Rose

"/X\(‘-‘)/X\"
Jan 2013

55618 Posts

Quote:
 Originally Posted by RichD Time to put an end to this nonsense. It was a nice ride, nearly 20 fold in stats. I get more from a free account than a paid. There is no free lunch. I'll continue my local GTX 560 Ti farm for a while.
Do you pay for electricity? I stopped running my GTX 580s because I no longer had free power, and the savings of a newer card would pay for the card quickly.

3 580s would cost me about $2.70/day to produce about 1300 GHz-days of TF. 1 2080 Ti would consume about$0.90/day to produce about 3500 GHz-days of TF. That's about 8 times as energy efficient.

If we increase the number of 580's to 8 to do the same work, that would cost me $7.27/day instead of$0.90/day. That's a savings of \$6.37/day. That buys a used RTX 2080 Ti in less than 4 months in my market. Buying a new RTX 3070 gives even more throughput using less power than a 2080 Ti, for about the same money.

It's something to consider.

2020-11-08, 17:36   #245
chalsall
If I May

"Chris Halsall"
Sep 2002

2×3×1,567 Posts

Quote:
 Originally Posted by Mark Rose It's something to consider.
It is indeed.

What's ***really*** weird about this situation is those on the Free Tier continue to get a reasonable amount of GPU compute. As an example, across my 13# Colab instances (each singularly tied to a Google Gmail account), I continue to reliably get ~2.5 hours of T4 allotment each and every day.

Meanwhile, I've been observing RichD, kladner, LOBES et al who I *know* are running the Paid Tier, and they are lucky to have an instance survive for more than a few minutes.

As usual with Google, even though these *customers* are paying (admittedly a very small amount) for the service, getting a response from a human representative is demonstrably almost impossible.

Sigh... Makes forcasting of assignment strategy even more difficult than it already is...

Last fiddled with by chalsall on 2020-11-08 at 17:37 Reason: Smelling mistakes.

2020-11-08, 19:55   #246
RichD

Sep 2008
Kansas

7×463 Posts

Quote:
 Originally Posted by Mark Rose Do you pay for electricity?
Currently not. It is all inclusive (utilities) in the rent. I'll continue running my two free accounts. When I complained to billing they said, "are you sure you are not running into the idle timeout?" Good grief, if they would look up my usage they would see I have been doing this for months.

 2020-11-08, 19:57 #247 lycorn     Sep 2002 Oeiras, Portugal 11·131 Posts I am still getting 12 hours on every CPU session I run, in general two concurrent sessions per gmail account. I have been running very few GPU sessions as of late, but every time I try one, I get a few hours of a T4 / P100, if I take a reasonable time between connections. All sessions are free. I wonder whether the location we are connecting from has some kind of influence (higher local demand leading to less resources provided).
2020-11-08, 20:00   #248
lycorn

Sep 2002
Oeiras, Portugal

11×131 Posts

Quote:
 Originally Posted by RichD When I complained to billing they said, "are you sure you are not running into the idle timeout?" .
Lol... as if there was such thing as "idle" in GIMPS.

2020-11-08, 20:55   #249
chalsall
If I May

"Chris Halsall"
Sep 2002

2×3×1,567 Posts

Quote:
 Originally Posted by lycorn Lol... as if there was such thing as "idle" in GIMPS.
LOL indeed!!!

We have actually encountered something like this once before (twice if you include Kaggle in the sample set)...

I don't have the time to drill down on exactly when, but several months ago no one was getting Colab GPU instances for more than a few minutes (if even offered them). Interestingly, CPU-only instances seem to have been always available.

I personally had written off Colab GPU somewhat along the lines of "Sincelerly, thanks for all the fish..." and then suddenly a couple of weeks later we could access it again.

I agree with kladner that this is likely a psychology / economic cross-over experiment for Google, and we're just somewhat extreme "edge-cases" they need to include in their model...

Thoughts? Observations?

2020-11-08, 21:31   #250
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

113558 Posts

Quote:
 Originally Posted by lycorn I am still getting 12 hours on every CPU session I run, in general two concurrent sessions per gmail account. I have been running very few GPU sessions as of late, but every time I try one, I get a few hours of a T4 / P100, if I take a reasonable time between connections. All sessions are free. I wonder whether the location we are connecting from has some kind of influence (higher local demand leading to less resources provided).
Count your blessings. Here in middle of USA, cpuonly may go 12 hours, or 12 minutes, or anything in between, especially low on weekdays. An account may not get a gpu at all for days, and a good day is when any gpu lasts 90 minutes. P100 are rare.

Quote:
 Originally Posted by RichD I'll continue running my two free accounts. When I complained to billing they said, "are you sure you are not running into the idle timeout?" Good grief, if they would look up my usage they would see I have been doing this for months.
My experience in other contexts (ISPs etc) is level 1 tech support is not capable of or perhaps not authorized for more than following a script (such as independent thought), and that seems to go for the rest of direct initial customer contact as well.

2020-11-08, 22:20   #251
Runtime Error

Sep 2017
USA

5×43 Posts

Quote:
 Originally Posted by chalsall What's ***really*** weird about this situation is those on the Free Tier continue to get a reasonable amount of GPU compute. As an example, across my 13# Colab instances (each singularly tied to a Google Gmail account), I continue to reliably get ~2.5 hours of T4 allotment each and every day.
I only run free colab, but I see the same of about ~2h30min if I run every day. If I run two sessions concurrently on the same gmail, I get half the time on each session (~1h15min) for the same total time. And if I miss a day or two, I tend to get double or triple the time the next run. So one session every other day gives ~5 hours of T4.

 2020-11-08, 22:35 #252 LOBES   Mar 2019 USA 568 Posts My account is paid up until November 29. If things improve to how they used to be (24hrs per single GPU session) then I'll renew. If not, I'll let it expire and just do a freebie account. I feel bad for chalsall, as I imagine throughput at GPU72 will go down a little.
 2020-11-09, 05:56 #253 DrobinsonPE   Aug 2020 10100102 Posts I have one free account that I try to keep going as much as possible. I just leave the tab open and hit the play button any time I see it stopped. I usually get one 2-3 hour session with a GPU and one 12 hour session with just a CPU each day. once or twice a week I catch it just right and get another 7-9 hour CPU session per day. Except for a two week period about a month ago, this has been my experience with the free account. The GPU sessions are about 50% P100, 30% T4, and 20% K80. I just go with whatever I get for the day. I agree with whoever said that this is a psychology / economic cross-over experiment for Google. About a month ago I tried some of the tricks everyone mentioned to dump the K80 sessions and fish for the T4 sessions. Right after I started doing that, my GPU session length dropped from 2-3 hour down to 45 minutes or less. About 1 week after I gave up on trying to always get a T4, my sessions went back to normal.