Colab question

[Uncwilly]

I have seen 2 or 3 cycles in the beginning. The GPU type get reported. It runs through again the GPU type gets reported etc. It does start the test after it has looped 1 or 2 times. (Don't ever think I have seen more.) Doesn't seem harmful, so I haven't said anything.

[chalsall]

Originally Posted by Uncwilly

Doesn't seem harmful, so I haven't said anything.

OK. Thanks for the report.

I don't understand why it happens, but *very* rarely the Notebook payload doesn't get work, and so the (newish) loop catches (and recovers from) this edge-case.

Having done some drill-down a while ago, I suspect that it's a temporary networking issue between the Colab instance and the GPU72 server. In cases where I've observed this myself, there's no record in the server's logs of an initial fetch attempt.

Concurrency is fun!

[kladner]

I got a T4 just now on my maiden voyage with Colab.
It looks like it is doing 1 class per minute.

[PhilF]

Originally Posted by kladner

I got a T4 just now on my maiden voyage with Colab.
It looks like it is doing 1 class per minute.

T4 on your first try? Sounds like a typical bait-and-switch strategy to me.

[kladner]

My expectation are limited. It's good while it lasts.
BTW: How often does Colab submit results, and where do they appear?

[chalsall]

Originally Posted by kladner

BTW: How often does Colab submit results, and where do they appear?

In order to automatically have results be submitted to Primenet, the system needs to know your Primenet Username (but not your PW). Please go to your account settings page, and enter this information on that form.

Then your results appear in your usual Completed Assignments report.

Thanks for trying out the GPU72_TF'ing thing. Also, so you know, when GPU72 knows your PNUN, you'll also be given P-1 work to do in parallel, on the CPU. Not terribly fast, but it all adds up...

[kladner]

Thanks! Would that be the Login or the Public name?
Nevermind. The mention of password not needed answers that.

Working on the 9th TF now.

Do I need to restart the runtime for this to take effect?

[chalsall]

Originally Posted by kladner

Do I need to restart the runtime for this to take effect?

Nope. The auto submission starts as soon as I do a bit of "magic" in the background, which I just did. You'll now see a GPU72_TF "Virtual Machine" in your Primenet Computer's list, and the eight just-submitted results.

The P-1 CPU work won't be assigned until you start your Notebook again. It's safe to stop and then restart your current running session, or you can just wait until your next run. Up to you.

[RichD]

Originally Posted by chalsall

Also, so you know, when GPU72 knows your PNUN, you'll also be given P-1 work to do in parallel, on the CPU.

I'm sure this has been considered before but the ones which have not added the PNUN (Primenet Username) would it be beneficial to sieve on the CPU and let the GPU be dedicated to TF?

[kladner]

The auto submission starts as soon as I do a bit of "magic" in the background, which I just did.

Woohoo! Thanks Chris.
I actually hate to restart while I have a T4 crunching, especially since that seems like a privileged assignment. After this many hours I'll probably have limits kicking in sometime.

EDIT: Update: The first run, with the T4 lasted many hours. I'm still not tuned in enough to see how long. I fumbled around this morning a good bit before I got the notebook running again. It came up with a P4, which seems to run at about 0.4 of the T4 speed. P-1 started up and reports every 5 minutes, interleaved with the TF output.

Another edit: Unlike Covid-19, where flattening the curve is the object, Colab is starting to bend my 30 average curve upward.

[chalsall]

Originally Posted by RichD

I'm sure this has been considered before but the ones which have not added the PNUN (Primenet Username) would it be beneficial to sieve on the CPU and let the GPU be dedicated to TF?

Actually, I hadn't thought of that... Although I suspect that with the CPUs so (relatively) underpowered, and the GPUs so high-end, that there wouldn't be much upside in doing that.

One thing I've modeled is to give P-1'ing work to those who's PNUN isn't known by GPU72, but under a collective Primenet account. Each worker would get the credit on GPU72, but not Primenet. (Assuming, of course, that the worker hasn't disabled CPU work in their instance(s) settings.)

Barbados is currently suffering hugely degraded international connectivity at the moment. FLOW, our provider, is either having OSPF and/or BGP issues. Frustrating!