There is no much performance difference after I enabled the MASTER_READTITLE_ONLY flag on the clusters I used. I changed it to default for future release in case other users have same problems on a particular system. There could be a little negative impact if the network between compute nodes is slow, but that should be minor. As this flag will control how many cores read the tile part simultaneously, it might not affect the performance very much for a small number of CPU cores. The performance of I/O operations for the system should be examined for this case.
Yusong
Is CWiggler element parallelizable
Moderators: cyao, michael_borland
Re: Is CWiggler element parallelizable
Yusong,
Thanks for all the help.
It might be simply caused by an incomplete build before I enabled the flag. I usually did the build for serial first and I might forgot to do make clean before make Pelegant.
I got it work under both cygwin in a Vista quad core desktop and Fedora in a quad core laptop. Now it only takes about 2 hours to finish the jobs which took 12 hours before.
Thanks,
Wanming
Thanks for all the help.
It might be simply caused by an incomplete build before I enabled the flag. I usually did the build for serial first and I might forgot to do make clean before make Pelegant.
I got it work under both cygwin in a Vista quad core desktop and Fedora in a quad core laptop. Now it only takes about 2 hours to finish the jobs which took 12 hours before.
Thanks,
Wanming
Re: Is CWiggler element parallelizable
Great to know you got super-linear speedup with Pelegant.
Yusong
Yusong