Google Diet Colab Notebook

[lycorn]

Ok, I don´t know why I was convinced each VM had two cores, but in the end it was good I was.
Otherwise I wouldn´t have tried running 2 NBs under the same account, at least for the same type of work, as HT would be pretty useless.
As I am using 5 free-tier accounts, I don´t think I´ll push my luck any further, as getting even 2 VMs per account not always works, let alone 3 or 4...
Oh well, all this for free, no complaints whatsoever .

[Ensigm]

Originally Posted by axn

Colab VMs don't have two cores - you just have one hyperthreaded core (i.e. 1c/2t). Therefore, you should use 1 worker 1 thread, and if you want to see if hyperthreading helps, enable the option of "use hyperthreading for LL" (please lookup the correct option for this setting), which will use both threads for a single test. That means you get the benefit for P-1 as well.

Good news, hyperthreading helps. Now I have this free 6~7% performance boost with no impact on memory usage.

[chalsall]

Originally Posted by Ensigm

Good news, hyperthreading helps. Now I have this free 6~7% performance boost with no impact on memory usage.

Really!!! I left ~6% of the free Google Colab P-1 compute on the table because I didn't think to test this??? Man... Never trust a human to do a job optimally...

Thanks, Ensigm. I will be sure to add this to my todo list to enable within the GPU72 Notebook for its next release (maybe in a week or so).

[axn]

Originally Posted by Uncwilly

I have never been able to get 2 to run in parallel on a single account (on purpose).

Interesting. Were you trying to run the same notebook from two different sessions or two different notebooks?

FWIW, I'm running different notebooks, and I'm also running them from a private window (to avoid my google account from interfering my other browsing). I am also at a location where paid tier is not available.

EDIT:- Could also be a recent development @ colab

EDIT2:- Could also be a CPU-only thing.

[Ensigm]

Also different cpus may have different speedups regarding to multithreading. For example, Model 63 (you can view this by executing !lscpu) can see a 10~15% improvement with multithreading enabled, while Model 85 gets probably only 5% or less. (That is, speaking only from my memory, as I didn't bother to stop my Stage 2 P-1 to benchmark it.) The speedup might also depend on the FFT size.

[axn]

Well, they put an end to that real quick, didn't they. No more parallel sessions. Now they've put in limits to only 1 session at a time.

[lycorn]

I´m still (00:10 UTC) running two NBs per account on 4 out of the 5 accounts I use.The fifth one was only allowed one. No changes here, then. Let´s see how long it will last.

[bayanne]

Even on free accounts now, very few GPUs are being allocated. Ah well, it was good whilst it lasted.

[De Wandelaar]

Originally Posted by bayanne

Even on free accounts now, very few GPUs are being allocated. Ah well, it was good whilst it lasted.

Indeed : no the least GPU allocated yesterday.

[Uncwilly]

I got >2:30hrs on T4's on each of 2 yesterday.

[LaurV]

Same like Unc, here we never managed two sessions per account. When we start the session at home, the one at job is killed. And viceversa. The started one continues from the stopped one left. They are different computers, different ISP, even different cities (yep, we commute for work to a nearby city). Never mixed. Except the gugu account is the same for this sole purpose (of using colab). (edit: yep, different keys from Chris, too).

We managed to run more than one session using different accounts. Personal, I mean. But the trouble is too much (or laziness too high), so we gave up, or only do this rarely.

SWMBO's account is different story.