From gregory@andros.physics.arizona.edu Fri Mar 12 13:53:46 2004 MIME-Version: 1.0 Date: Fri, 12 Mar 2004 11:53:44 -0700 (MST) From: Eric Gregory To: brower@super.bu.edu cc: brower@lns.mit.edu Subject: MILC on QCDOC benchmarks X-IMAPbase: 1208291383 0000000002 X-UID: 1 Status: O Content-Length: 1674 Bob, Here are some benchmarks that I have so far. They should be considered preliminary. Some are pretty short runs, but the numbers dont fluctuate much. These runs are for regular MILC (no L3 inverter) compiled with gcc. The two tables below are for double and single precision. For small local lattice sizes (4^4 for double, 4^4 & 6^4 for single) I tried putting the lattice data in the embedded edram. These numbers are marked with an *. We are still doing tests to make sure we are doing this in the best, most optimal way, but it seems to speed things up. Also it should be reiterated that with the 2^4 nodes partition, the communication in the plus direction and that in the minus directions are going over the same wire to the other node in that dimension. When a larger partition is available (hopefult at least 4^4 nodes) is available we should see some speedup in the per node performance, due to the communication being a little faster. Our plan for the next week or so before the all-hands meeting is to see if we can get the L3 inverter bolted into MILC in some way and start to get some numbers with wich to compare this. Also, as soon as I hear that xlc is available, I will recompile this with xlc for further comparision. I will also fill in this table with more partition and lattice geometries andf sizes. double L=4 6 8 10 12 nodes 1 51.0/67.1* 51.3 50.3 50.3 51.0 2^4 = 16 36.0/43.6* 43.7 45.7 47.8 single nodes L=4 6 8 10 12 1 56.6/75.8* 57.2/76.6* 56.0 57.3 56.7 2^4 = 16 41.3/50.5* 52.7/65.0* 52.7 54.4 55.0 L is the linear size of the local lattice. local lattice volume is L^4. -Eric From rdm@physics.columbia.edu Thu Mar 11 17:12:32 2004 Date: Thu, 11 Mar 2004 17:11:43 -0500 (EST) From: rdm@physics.columbia.edu To: brower@bu.edu Cc: chulwoo@physics.columbia.edu, nhc@physics.columbia.edu, pab@physics.columbia.edu, petrov@physics.columbia.edu Subject: performance figures for SciDAC meeting X-Score-Level: 0.285 X-UID: 2 Status: O Dear Rich, We at Columbia are working to have results at the all-hands meeting for the following cases simulator 128 node 128 node Fermion V D-slash D-slash CG ------- --- --------- --------- -- Wilson 2^4 47% X X 4^4 53% X X 8^4 X X Asqtad 2^4 X X 4^4 43% X X 8^4 X X The X's are the points we are working to fill in. Most of these cases have already been run as part of our ASIC verification. However, the code was not completely tuned up. In particular, we did not run the CG on the simulator, since it takes too long. We've had to optimize some of the CG's linear algebra, that doesn't involved D-slash, since the D-slashes are so fast. Best, Bob