Hi Tosa, There should be no disk swapping at all. All the machines have 6G of RAM. Even more than a single 32-bit process can handle. So disk shouldn't be a factor.
Regarding P4s. Yes, certain aspects of it are certainly slower than P3s, but some of those problems do not exist in the P4 Xeons, which is what these machines are. And for various other computational tasks I have seen a 3 fold decrease in time. A recent benchmark provided a comparison of the Pentium-M 1.6GHz to the P4 (not a Xeon) 2.6GHz as being roughly similar. But relative to the P3-1GHz (which does not have a huge L2 cache), the PM 1.6 is vastly superior (primarily because of its 1M L2 cache). We are running 2.8G P4 Xeons and I don't care about hyperthreading because Linux didn't for the 2.4 kernels. It is a feature of the 2.6 kernel and does provide a 50% speed improvement, but we haven't changed our machines yet. But back to the original question, why doesn't a P4 Xeon 2.8G processor provide at least one and a half times the speed benefit of a 1GHz P3. I am not asking for a 2.8x speed increase. Thanks, Satra -----Original Message----- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] On Behalf Of Yasunari Tosa Sent: Wednesday, January 14, 2004 2:31 PM To: [EMAIL PROTECTED] Subject: Re: Computation times The most basic reason ================ Unfortunately Freesurfer programs are very memory hungry. Once swapping starts, the time is disk I/O bounds. The processor speed won't matter much. Are you sure that your runtime won't have any swapping activities? Next item ======= You cannot compare P4 vs. P3. The P4 processor design was crippled at the end to make money quick and out to market (since P4 is much cheaper to make than P3 and additional features made more expensive and more time to verify). PM (up to 1.6 Ghz, updated P3) can beat P4 3 GHz processor in performance. AMD Athlon XP and Opteron processors below 2 Ghz can beat 3 GHz P4. What you should do is to compare within the P4 processors. I'm sure that Xeon is faster than P4 at the same clock, because of bigger cache. Third item ======== Freesurfer is not using threads at all and thus hyperthreading Xeon and P4 architecture won't help. Actually I found that there are many locations we could use pthreads very easily. You may start seeing threads soon. Tosa Satrajit Ghosh wrote: >Hi Bruce, > >I may have asked this question before, but some part of the issue still >doesn't make sense and hence the need for some clarification. > >According to recon-all stages 4a and b should take 7.5hrs per >hemisphere on a P3 1GHz machine. We have a P4 Xeon dual 2.8 GHz with >enough memory and it takes somewhere between 6 and 7 hrs. Since the >topology has already been fixed, shouldn't the timing of these stages >scale inversely with processor speed? Are there other factors which influence the timing of these stages? > >Or a different way to put it, what should we do to get a speedup factor >that a faster machine provides? > >A possibly related question, does freesurfer use operations on NaNs for >any stage of computation? > >Thanks, > >Satra > > >
