HPSS access from PDSF

["Craig E. Tull" <cetull@lbl.gov>]

HPSS-Users;

I have recently finished some tests of the network connectivity between the
PDSF room and the NERSC Mass Sortage system (HPSS). These tests have been
conducted using the PDSF STAR cluster (starsu00, starlx01-08) since they
have 100bT connections to the PDSF network switch and are fast enough to
give reasonable results.

SOLARIS <<<<<<<<<<<<<<<<<<

starsu00 -> HPSS COS1 : write using pftp/pput
	2.12 MByte/sec

starsu00 <- HPSS COS1 : read using pftp/pput
	2.25 MByte/sec

starsu00 -> HPSS COS3 : write using pftp/pput/explicit COS
	4.07 MByte/sec

starsu00 <- HPSS COS3 : read using pftp/pget
	5.75 MByte/sec

LINUX <<<<<<<<<<<<<<<<<<<<<

starlx06 <- HPSS COS3 : read using ftp/get
	4.40 MByte/sec

The lesson here is this:

The large Class Of Service (COS1) is much slower than the medium Class Of
Service (COS3) for HENP-like access. After speaking with Nancy Meyer, I
believe the reason is two-fold:

1> The connection between PDSF and the HPSS goes through a 10 bT connection
to access COS1.
2> The HIPPI mounted MaxStrat disk cache is very heavily used and can form
a bottle neck (ie. while your job waits in a queue with Cray jobs).

- Craig


-----------------------------------------------------
 Craig E. Tull, Ph.D.
 National Energy Research Scientific Computing Center
 E.O. Lawrence Berkeley National Laboratory
 EMail: cetull@lbl.gov       WWW: http://iago.lbl.gov/
 Phone: (510)486-7253        FAX: (510)486-4004