DPUs on DDR: In-memory processing

[kriesel]

Originally Posted by ewmayer

Yes, one may have the issue of working set being larger than the obvious alloc'ed stuff in the program. I do know that Prime95 uses larger aux-data arrays than Mlucas, ... even so, some code tweakage might be needed to get the working set size to fit, since IIRC that includes more than just the explicit code-allocated stuff, e.g. libraries.

Quite so; I happen to have Mlucas V17.0 running a 100Mdigit exponent on one Intel core and it occupies 197M working set, much less per exponent bit than the prime95 example I gave. Probably the prime95/mprime GUI vs. console difference is small.
224M/79.34M = 2.82
197M/332.2M = 0.593
ratio 2.82/0.594 = 4.76.
No complaints, just an observation, and comparing oranges to watermelons.

[Prime95]

This surprises me. I would expect prime95 to have a working set size of about 45MB for a 4M FFT. Does mprime have a lower working set than prime95?

[kriesel]

Originally Posted by Prime95

This surprises me. I would expect prime95 to have a working set size of about 45MB for a 4M FFT. Does mprime have a lower working set than prime95?

No data on mprime currently.

Memory indications are from Windows task manager in all cases below.

I stopped, exited, and restarted the 79.34M prime95 29.8b3 1-worker instance; it shows 221MB private working set, 229MB peak working set; it's a Win7 x64 system. (parrot)

Different system (falcon), Win 10 x64, prime95 29.8b5, 2 workers, 86.3M PRP in each, 345MB memory

Third system (condor), Win 7 x64, prime95 29.8b3, 6 workers, 5x ~86M PRP & one P-1 limited to 8192MB currently in stage 1 (stage 1 does not indicate memory usage in worker window title bar), 1309MB current private working set, 8977MB peak working set.

Fourth system (condorella), Win 7 x64, prime95 29.8b3, 4 workers, 3x ~86.3M PRP & one P-1 currently in stage 2 indicates using 7851MB itself, 8811 MB private working set

[chris2be8]

One of the uses they list is sparse matrix multiplication (in slide 8 on https://www.anandtech.com/show/14750...ing-by-upmem/3 ). Would this help linear algebra (msieve or CADO-NFS)?

Chris

[ewmayer]

Originally Posted by chris2be8

One of the uses they list is sparse matrix multiplication (in slide 8 on https://www.anandtech.com/show/14750...ing-by-upmem/3 ). Would this help linear algebra (msieve or CADO-NFS)?

Chris

Since the In-memory DPUs are based on simple bitwise operations, I would think them ideally suited for NFS-style linear algebra.