[Bioclusters] RE: About the parallel Blast on PVM

[Joseph Landman]

Hi Filatei:

  There is a solaris64.ncbi.mk target for the NCBI blast.  It is
possible that your version was built using one of the other tools.  If
you want to build it for the 64 bit Solaris architecture, you should
grab the NCBI source code, run the make script, and test the resulting
binaries.  

  The NCBI source is at ftp://ftp.ncbi.nlm.nih.gov in the tool_box
directory (site is down right now, so I cannot give specifics).  Do the
following on the target machine.

1) Uncompress it, untar it.  

2) Edit the ncbi/platform/solaris64.ncbi.mk to reflect any option
changes you wish, specifically looking at the following flags:

    NCBI_CC = cc -xtarget=ultra -xarch=v9 -xildoff
    NCBI_CFLAGS1 = -c -Xa
    NCBI_LDFLAGS1 = -Xa
    NCBI_OPTFLAG = -fast -xtarget=ultra -xarch=v9

You may wish to change the options, specifically NCBI_OPTFLAG and
NCBI_CFLAGS1.  

3) Build the system:

	ncbi/make/makedis.csh

After some time (and hopefully without errors), you should have a few
new binaries in the ncbi/build/ directory.  You should be able to
execute

	ncbi/build/blastall

and see all the blastall arguments.

If this doesn't work, call you local IT/support person to give you a
hand.  This procedure should generate the 64 bit binaries on a 64 bit
Solaris platform (provided the scripts recognize it as such).

Please note that there are reasons why the 64 bit code is bound to be
slightly (to possibly significantly) slower on smaller (e.g. less than 2
GB) data sets.  Pointer and integer sizes are 64 bit on a true 64 bit
platform, which means that fewer can fit in cache (which lowers the
effective size of the cache by decreasing the data stored within it),
and requires more time to transfer than a 32 bit word (which reduces the
data rate to the CPU, as the CPU can grab only half as many 64 bit
variables as 32 bit variables at a time). 

The net effect for data sets less than 2 GB in size is a speed
reduction.  Couple this with the speed and price advantage of the IA32
platforms (Athlon and P III) on BLAST as compared to RISC platforms, and
it might not be worth the effort to rebuild BLAST for the Sun if you
already have a few well configured late model PC servers about.

Joe


On Mon, 2002-08-19 at 19:08, Akpodigha Filatei wrote:
> Hi, I am interested in this as well.  I have NCBI blast running twice faster on a 2-processor Intel Linux box than 4-processor UltraSparc server(SUN E3500, Solaris 8).  It is not using more than 1 processor on the solaris server even when we specify 4 or more for the -a parameter, as it does on the Linux server.  Is there any known implementation that takes advantange of the 64-bit architecture of Sun servers?  Please let me know if you have any useful info on this.
> 
> Filatei 

-- 
Joseph Landman, Ph.D
Scalable Informatics LLC
email: landman@scalableinformatics.com
  web: http://scalableinformatics.com
phone: +1 734 612 4615