[Bioclusters] Clusters

[Michael Hanulec]

Juha these are some very good tests that I wish I could perform myself,
but I have no access to Apple G5 hardware.

> We had high hopes for Apple G5 after reading the blast/hmmer etc.
> comparisons on Apple www-pages. Apple kind borrowed us a dual-G5 2 GHz
> with 2 GB RAM and fast disks box. We ran different blasts and hmmer
> searches, both with binaries we compiled (gcc) and precompiled from
> Apple's ftp ("AG-blast") and hmmer www-page (which to best of our
> understanding are the same Apple refers in their ads). In addition we
> ran some statistical genetics apps (merlin and pseudomarker). We were in
> contact with Apple (actually a very knowledgeable tech guy form Apple!,
> but they couldn't really explain the differences we observed).

I'd be interested what would of happend if the binaries were compiled with
the IBM XL C/C++ and Fortran compilers:

http://www.absoft.com/Products/Compilers/C_C%2B%2B/XLC/XLC.html
http://www.absoft.com/Products/Compilers/Fortran/Macintosh/XLF/xlf.html

Unfortunately not all version of GCC are created equal.  I doubt the
opimization quality of the code produced by the OS X and x86-64 versions
is equal.

> The Opterons were IBMs e325 boxes with SCSI RAID-1 disks. Concerning
> blast & hmmer types of runs, we thought this could be explained by disk
> access speed (G5 had SATA drive), therefore we also mounted the blast /
> hmmer dbs over NFS (1 GBit Ethernet) on the Opterons, but this really
> did not change anything (1st runs were slower than from internal SCSI
> disks, but we produced the numbers as averages from runs 2-4). Both CPUs
> were in full use in G5s and Opterons, memory was not limiting factor (ie
> swapping etc. did not occur) and power-saving features etc. were
> switched off. The only type of app Altivec-enhanced G5 performed better
> than Opterons was "AGBlast" with wordsize 40 (but not with typical
> smaller word size). (Note also that there have been discussions about
> whether the performance increase in the "AGBlast" is from AltiVec or
> just from other non-platform specific optimization of the original NCBI
> blast code; the same optimizations are available for Intel/AMDs, but we
> did not try those). In addition, all non-AltiVec optimized apps
> performed slower than Opterons (no code optimizations were done in any
> tests, except with the AltiVec).

[ Just in case not everyone knows this MSI makes this Opteron server for
IBM.  The model is called the K1-1000.  ]

Again.. the use of commerical compilers to run these test would be
interesting also.  Currently there are a choice of three commerical AMD64
compilers: Absoft, PGI and PathScale.

-mike
--
hanulec@hanulec.com		cell: 858.518.2647 && 516.410.4478
https://secure.hanulec.com	      EFnet irc && aol im: hanulec