GMP-ECM GPU for WSL problems

[Rubiksmath]

Originally Posted by SethTro

would you mind sharing the result of running?

$ echo "10^248+123" | ./ecm -v -cgbn 2e5 0

Okay, here it is:

CGBN<1024, 8> running kernel<56 block x 256 threads> input number is 824 bits
Computing 100 bits/call, 1/288578 (0.0%)
Computing 110 bits/call, 101/288578 (0.0%)
Computing 121 bits/call, 211/288578 (0.1%)
Computing 256 bits/call, 1585/288578 (0.5%)
Computing 655 bits/call, 5631/288578 (2.0%)
Computing 2997 bits/call, 220921/288578 (76.6%), ETA 2 + 6 = 8 seconds (~5 ms/curves)
Copying results back to CPU ...
Computing 1792 Step 1 took 8ms of CPU time / 8414ms of GPU time
Throughput: 212.971 curves per second (on average 4.70ms per Step 1)

Also, yeah I don't fully know what happened with the WSL, in the end I just installed Ubuntu dual boot and did it that way and it worked. I have a feeling its a problem on my end with how I've installed the drivers, I can try and see if I can install it on WSL again.

If it helps, here is the last line output from running configure (although as I said I've probably scuffed the drivers pretty bad so I'll try this later on a fresh install)

./configure --enable-gpu --with-cuda=/usr/local/cuda
......
checking for cInit in -lcuda ..no
configure : error: Couldn't find cuda lib

depending on how I had set it up I did get other errors, but all of them had something to do with the CUDA toolkit, as I said probably my bad and I'll try it again soon to see if I can work out what went wrong.

[wombatman]

Originally Posted by SethTro

I'm glad it's much faster! I'm curious how my 1080ti stacks up against the 30 series (if someone in this thread has a 3080/3090 I'd love to see some timing results too), would you mind sharing the result of running?

$ echo "10^248+123" | ./ecm -v -cgbn 2e5 0
...
CGBN<1024, 8> running kernel<56 block x 256 threads> input number is 824 bits
...
Computing 1694 bits/call, 134631/288578 (46.7%), ETA 8 + 7 = 14 seconds (~8 ms/curves)
Copying results back to CPU ...
Computing 1792 Step 1 took 37ms of CPU time / 14013ms of GPU time
Throughput: 127.885 curves per second (on average 7.82ms per Step 1)

With a 3090:

CGBN<1024, 8> running kernel<164 block x 256 threads> input number is 824 bits
Computing 100 bits/call, 1/288578 (0.0%)
Computing 110 bits/call, 101/288578 (0.0%)
Computing 121 bits/call, 211/288578 (0.1%)
Computing 256 bits/call, 1585/288578 (0.5%)
Computing 655 bits/call, 5631/288578 (2.0%)
Computing 2725 bits/call, 203513/288578 (70.5%), ETA 3 + 6 = 9 seconds (~2 ms/curves)
Copying results back to CPU ...
Computing 5248 Step 1 took 2393ms of CPU time / 8824ms of GPU time
Throughput: 594.753 curves per second (on average 1.68ms per Step 1)

[SethTro]

Originally Posted by wombatman

With a 3090:
Computing 5248 Step 1 took 2393ms of CPU time / 8824ms of GPU time
Throughput: 594.753 curves per second (on average 1.68ms per Step 1)

Thanks for the benchmark. It's takes 35% less time AND does 3 times an many curves at a time, that's a big improvement! Makes me want to upgrade my system :)

I occasionally look at this big run, 4.3e9 x 100,000 curves for M1217, which would complete in batches of 1792 every 3.5 days over 200 total days on my 1080. My mind races when I think about that being batches of 5248 every 2.2 days with a 3090.

I wonder why it's taking 2.3seconds of CPU time. On my system CPU time is generally very small (sub 100ms). If you could run "$ echo "10^248+123" | time ./ecm -cgbn 1e5 0" and share the timing results (e.g. "0.04user 0.10system 0:07.21elapsed 2%CPU") and any anecdotes about it taking time to print the first "Computing 100 bits/call" line (or after "copying results back to CPU").

I think the CPU time is overlapped with the GPU time but it's possible that they aren't in which case there's maybe a 25% speed up.

[wombatman]

Originally Posted by SethTro

Thanks for the benchmark. It's takes 35% less time AND does 3 times an many curves at a time, that's a big improvement! Makes me want to upgrade my system :)

I occasionally look at this big run, 4.3e9 x 100,000 curves for M1217, which would complete in batches of 1792 every 3.5 days over 200 total days on my 1080. My mind races when I think about that being batches of 5248 every 2.2 days with a 3090.

I wonder why it's taking 2.3seconds of CPU time. On my system CPU time is generally very small (sub 100ms). If you could run "$ echo "10^248+123" | time ./ecm -cgbn 1e5 0" and share the timing results (e.g. "0.04user 0.10system 0:07.21elapsed 2%CPU") and any anecdotes about it taking time to print the first "Computing 100 bits/call" line (or after "copying results back to CPU").

I think the CPU time is overlapped with the GPU time but it's possible that they aren't in which case there's maybe a 25% speed up.

Oh, the CPU time is probably higher because I'm running other tasks as well (but wasn't on the GPU) that are using a good chunk of CPU time.

Running your requested command line gives:

GMP-ECM 7.0.5-dev [configured with GMP 6.2.1, --enable-asm-redc, --enable-gpu, --enable-assert] [ECM]
Input number is 10^248+123 (249 digits)
Using B1=100000, B2=0, sigma=3:2295140881-3:2295146128 (5248 curves)
GPU: Using device code targeted for architecture compile_86
GPU: Ptx version is 86
GPU: maxThreadsPerBlock = 896
GPU: numRegsPerThread = 67 sharedMemPerBlock = 0 bytes
Computing 5248 Step 1 took 1490ms of CPU time / 4739ms of GPU time
1.65user 3.56system 0:05.63elapsed 92%CPU (0avgtext+0avgdata 126448maxresident)k
20288inputs+0outputs (129major+3901minor)pagefaults 0swaps