BOINC NFS sieving - NFS@Home

[VictordeHolland]

GC_2_795

prp88 factor: 1229181211256783532456105424311951957003782336757922948778898041952727344588487561131449
prp122 factor: 29437568931055869616663350030336306362482066093333700749117926399399022787021904381727616483972300349043385650188711287347

95.0 hours with a "target_density=120" 10.6M matrix on a i7 3770k -t 4

[swellman]

prp84 factor: 148902775642108616331795830640649037143571534869122755521883005939595673614910241763
prp113 factor: 12971143073736097580217855171017913040659435287994287740254244481861979067885275717801474448610862951431512611037

[VictordeHolland]

LA started on L1282, ETA October 1st.

Couldn't build a matrix at target density 120 (not enough relations), so settled for a 20.3M matrix with target density 100.

Edit: I just noticed NFS@home increased the Q sieving limit of L1282, so I'll wait for the extra relations and build a new matrix.

[wombatman]

GC_5_353 factors as:

Tue Sep 16 00:48:21 2014  prp57 factor: 485969537734693364126687271492949711480027245728307842357
Tue Sep 16 00:48:21 2014  prp190 factor: 2236662401843371642445552346818808997187163195483097065186825555807216271969665340930331577121304317764383946464160253639913952177421808512066356909004894530314082698485713276741002939780381

Took 145 hours on i7-2630QM with 7 threads.

[pinhodecarlos]

Originally Posted by wombatman

GC_5_353 factors as:

Tue Sep 16 00:48:21 2014  prp57 factor: 485969537734693364126687271492949711480027245728307842357
Tue Sep 16 00:48:21 2014  prp190 factor: 2236662401843371642445552346818808997187163195483097065186825555807216271969665340930331577121304317764383946464160253639913952177421808512066356909004894530314082698485713276741002939780381

Took 145 hours on i7-2630QM with 7 threads.

On your laptop it is quicker to run on only 3 or 4 threads. Turn off HT. I think you just wasted 10-15 % of your total time.

[wombatman]

Originally Posted by pinhodecarlos

On your laptop it is quicker to run on only 3 or 4 threads. Turn off HT. I think you just wasted 10-15 % of your total time.

I've seen this mentioned before. Why does using fewer threads without hyperthreading help?

[xilman]

Originally Posted by wombatman

I've seen this mentioned before. Why does using fewer threads without hyperthreading help?

Think: is your process compute bound or memory bound? How well do the memory accesses fit in the caches?

[pinhodecarlos]

I think it is a problem on how LA phase is coded not supporting HT. Also with HT off the CPU will be cooler.

[wombatman]

Originally Posted by xilman

Think: is your process compute bound or memory bound? How well do the memory accesses fit in the caches?

Truthfully, I'm not sure. The 2630QM has 6 MB L3 cache, which from reading around, only has a strong effect if the data is being read sequentially. My assumption (based on my poor understanding of the black box workings) is that the LA step goes through some kind of sequential order. So would using 7 threads essentially saturate the L3 cache, making it the limiting step?

Am I in the ballpark?

[debrouxl]

Let's consider two other workloads for which I got benchmark data, before considering msieve:
* application A uses OpenMP for near-linear speedup on a loop with a huge number of iterations, each core chews through several dozen thousand iterations per second. All of the dataset fits in the cache, zero external memory accesses are performed after the initial load, and no floating-point operations are used.
The fastest computer (among the platforms I have access to) for that compute-bound workload uses a FX-8150 @ 3.6 GHz, which is a real 8-core system without HT, with a small L1 cache and only 4 FPUs, so it sucks at Prime95 LL testing, for instance. A couple Core-i7 HT Xeons @ 3.2 and 3.3 GHz are nearly as fast. A quad-core Cortex-A9 @ 1.7 GHz is less than 5 times slower than the FX-8150, so it's in the same ballpark.

* application B is single-threaded. Computer 2 has a CPU with a clock frequency more than twice higher than that of computer 1, definitely more cache and possibly a slightly newer micro-architecture (Xeon E5-1xxx vs. first-generation mobile Core i7), and uses DDR3-1600 while computer 1 uses DDR3-1333.
The workload runs less than 50% faster on computer 2 than on computer 1, so the workload is rather memory-bound. Using DDR3-2400, or the upcoming DDR4, would presumably yield near-linear speedup, but using a latest-generation Xeon at ~4 GHz clock speed wouldn't help that much.


Experience running msieve on an otherwise idle computer shows that:
* using 1 < N <= [real core count] threads decreases run time almost linearly;
* using [real core count] < N <= [hyperthread count] threads does at best slightly decrease the total run time, at worse increases it slightly, so the efficiency (runtime / number of threads) decreases near-linearly.
So msieve is a memory-bound workload indeed, and you should be using only 3 or 4 threads on your i7, as hinted by Carlos

[wombatman]

Thanks Lionel for the more detailed explanation. I guess this means I can cut back on the number of threads I use on my desktop as well, since I only have DDR3-2133 RAM there.