George found a performance bug with large messages in the RC. I was able to verify it. Netmpi numbers: 1.2.1rc1: 147: 134217730 bytes 3 times --> 11117.75 Mbps in 0.092104991 sec 148: 201326588 bytes 3 times --> 11108.94 Mbps in 0.138267001 sec 149: 201326591 bytes 3 times --> 11107.05 Mbps in 0.138290485 sec 150: 201326594 bytes 3 times --> 11116.16 Mbps in 0.138177156 sec "older" version (circa 2007): 147: 134217730 bytes 3 times --> 15962.75 Mbps in 0.064149340 sec 148: 201326588 bytes 3 times --> 15951.12 Mbps in 0.096294165 sec 149: 201326591 bytes 3 times --> 15962.12 Mbps in 0.096227845 sec 150: 201326594 bytes 3 times --> 15962.62 Mbps in 0.096224825 sec 151: 268435452 bytes 3 times --> 15946.49 Mbps in 0.128429532 sec 152: 268435455 bytes 3 times --> 15955.83 Mbps in 0.128354351 sec I'm going to bisect to see when this crept in. -d
Hmm. I went through the releases all the way back to 1.0.6p1, and I kept getting the bad numbers, so I have to figure out how I got the "older" numbers. -d On 11/13/2009 09:07 AM, Darius Buntinas wrote:
George found a performance bug with large messages in the RC. I was able to verify it. Netmpi numbers:
1.2.1rc1:
147: 134217730 bytes 3 times --> 11117.75 Mbps in 0.092104991 sec 148: 201326588 bytes 3 times --> 11108.94 Mbps in 0.138267001 sec 149: 201326591 bytes 3 times --> 11107.05 Mbps in 0.138290485 sec 150: 201326594 bytes 3 times --> 11116.16 Mbps in 0.138177156 sec
"older" version (circa 2007):
147: 134217730 bytes 3 times --> 15962.75 Mbps in 0.064149340 sec 148: 201326588 bytes 3 times --> 15951.12 Mbps in 0.096294165 sec 149: 201326591 bytes 3 times --> 15962.12 Mbps in 0.096227845 sec 150: 201326594 bytes 3 times --> 15962.62 Mbps in 0.096224825 sec 151: 268435452 bytes 3 times --> 15946.49 Mbps in 0.128429532 sec 152: 268435455 bytes 3 times --> 15955.83 Mbps in 0.128354351 sec
I'm going to bisect to see when this crept in.
-d
Did something outside of MPICH2 change? The compiler/glibc version or something like that? -Dave On Nov 13, 2009, at 10:23 AM, Darius Buntinas wrote:
Hmm. I went through the releases all the way back to 1.0.6p1, and I kept getting the bad numbers, so I have to figure out how I got the "older" numbers.
-d
On 11/13/2009 09:07 AM, Darius Buntinas wrote:
George found a performance bug with large messages in the RC. I was able to verify it. Netmpi numbers:
1.2.1rc1:
147: 134217730 bytes 3 times --> 11117.75 Mbps in 0.092104991 sec 148: 201326588 bytes 3 times --> 11108.94 Mbps in 0.138267001 sec 149: 201326591 bytes 3 times --> 11107.05 Mbps in 0.138290485 sec 150: 201326594 bytes 3 times --> 11116.16 Mbps in 0.138177156 sec
"older" version (circa 2007):
147: 134217730 bytes 3 times --> 15962.75 Mbps in 0.064149340 sec 148: 201326588 bytes 3 times --> 15951.12 Mbps in 0.096294165 sec 149: 201326591 bytes 3 times --> 15962.12 Mbps in 0.096227845 sec 150: 201326594 bytes 3 times --> 15962.62 Mbps in 0.096224825 sec 151: 268435452 bytes 3 times --> 15946.49 Mbps in 0.128429532 sec 152: 268435455 bytes 3 times --> 15955.83 Mbps in 0.128354351 sec
I'm going to bisect to see when this crept in.
-d
Maybe. This was on intel loaner (where I got the older numbers). A sanity check would be to compare to open mpi. -d On 11/13/2009 10:25 AM, Dave Goodell wrote:
Did something outside of MPICH2 change? The compiler/glibc version or something like that?
-Dave
On Nov 13, 2009, at 10:23 AM, Darius Buntinas wrote:
Hmm. I went through the releases all the way back to 1.0.6p1, and I kept getting the bad numbers, so I have to figure out how I got the "older" numbers.
-d
On 11/13/2009 09:07 AM, Darius Buntinas wrote:
George found a performance bug with large messages in the RC. I was able to verify it. Netmpi numbers:
1.2.1rc1:
147: 134217730 bytes 3 times --> 11117.75 Mbps in 0.092104991 sec 148: 201326588 bytes 3 times --> 11108.94 Mbps in 0.138267001 sec 149: 201326591 bytes 3 times --> 11107.05 Mbps in 0.138290485 sec 150: 201326594 bytes 3 times --> 11116.16 Mbps in 0.138177156 sec
"older" version (circa 2007):
147: 134217730 bytes 3 times --> 15962.75 Mbps in 0.064149340 sec 148: 201326588 bytes 3 times --> 15951.12 Mbps in 0.096294165 sec 149: 201326591 bytes 3 times --> 15962.12 Mbps in 0.096227845 sec 150: 201326594 bytes 3 times --> 15962.62 Mbps in 0.096224825 sec 151: 268435452 bytes 3 times --> 15946.49 Mbps in 0.128429532 sec 152: 268435455 bytes 3 times --> 15955.83 Mbps in 0.128354351 sec
I'm going to bisect to see when this crept in.
-d
OK, open mpi has similar low numbers. -d On 11/13/2009 10:26 AM, Darius Buntinas wrote:
Maybe. This was on intel loaner (where I got the older numbers). A sanity check would be to compare to open mpi.
-d
On 11/13/2009 10:25 AM, Dave Goodell wrote:
Did something outside of MPICH2 change? The compiler/glibc version or something like that?
-Dave
On Nov 13, 2009, at 10:23 AM, Darius Buntinas wrote:
Hmm. I went through the releases all the way back to 1.0.6p1, and I kept getting the bad numbers, so I have to figure out how I got the "older" numbers.
-d
On 11/13/2009 09:07 AM, Darius Buntinas wrote:
George found a performance bug with large messages in the RC. I was able to verify it. Netmpi numbers:
1.2.1rc1:
147: 134217730 bytes 3 times --> 11117.75 Mbps in 0.092104991 sec 148: 201326588 bytes 3 times --> 11108.94 Mbps in 0.138267001 sec 149: 201326591 bytes 3 times --> 11107.05 Mbps in 0.138290485 sec 150: 201326594 bytes 3 times --> 11116.16 Mbps in 0.138177156 sec
"older" version (circa 2007):
147: 134217730 bytes 3 times --> 15962.75 Mbps in 0.064149340 sec 148: 201326588 bytes 3 times --> 15951.12 Mbps in 0.096294165 sec 149: 201326591 bytes 3 times --> 15962.12 Mbps in 0.096227845 sec 150: 201326594 bytes 3 times --> 15962.62 Mbps in 0.096224825 sec 151: 268435452 bytes 3 times --> 15946.49 Mbps in 0.128429532 sec 152: 268435455 bytes 3 times --> 15955.83 Mbps in 0.128354351 sec
I'm going to bisect to see when this crept in.
-d
participants (2)
-
Darius Buntinas -
Dave Goodell