View Single Post
Old 2011-05-20, 09:27   #1
Karl M Johnson
Karl M Johnson's Avatar
Mar 2010

3×137 Posts
Default To CPU or to GPU: stage1 polyselect research

As the topic suggests, I did this little research on stage one of polynomial selection. Jason suggests doing only stage 1 on GPU in the latest Msieve NFS manual, instead of combined search.

You will find all the data in the table below(which is actually a picture).

Now, here are the numbers used:
c90: 69547303316492118149408022547436991785350544259733986981963408781389041467965565432357913
c100: 776701169123392332156259995503621920821842983186779441829856164218194104805531020576372063083759743
c110: 9852238794222235462151171006209225112694534204051316346901796414666349372723740059709872068581469631350831777
c120: 227026401879198463827454062763506967433946202931364283264305901265796403304094997028726671231101499391384899872706432273
c130: 1142067639655668244421675659299472228892561443039817892634518492620626571274348874167990804267595223355216327568784472374242262183
c140: 2333435938791497921844814519438006187812846866005520174850758490445958886619248051390445682925499819800889470401806186619479812725646985161
And the ranges for all of them:
1. 1,10000
2. 1000000,1010000
3. 1000000000,1000010000
Now, apart from measuring the running time, I also measured how much strings(lines) were produced by searching all 3 ranges on a certain number, for both GPU and CPU, in msieve.dat.m file. The more the better. I did a sort -u before counting.

Some specs:
CPU: Core 2 Quad Q6600 @ 3.0 Ghz
GPU: Nvidia GeForce GTX 480 @ 800/950 Mhz with 270.61 ForceWare
OS: W7 x64 with SP1
Msieve versions used:
CPU - 64 bit binary from latest SVN(thanks Jeff!)
GPU - 64 bit CUDA-enabled binary from latest SVN(thanks Brian!)
I used timethis from windows 2000 resource kit to measure the running time of binaries for each range.

Some conclusions:
1. On C90 and C100, CUDA-enabled Msieve performs better than the one which uses the CPU only.
2. Using some kind of script to utilize all available hardware would be the best way to do stage1 of polynomial selection(sort of obvious).
Attached Thumbnails
Click image for larger version

Name:	research.png
Views:	206
Size:	31.7 KB
ID:	6621  
Karl M Johnson is offline   Reply With Quote