[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference turris::digital_unix

Title:	DIGITAL UNIX(FORMERLY KNOWN AS DEC OSF/1)
Notice:	Welcome to the Digital UNIX Conference
Moderator:	SMURF::DENHAM

Created:	Thu Mar 16 1995
Last Modified:	Fri Jun 06 1997
Last Successful Update:	Fri Jun 06 1997
Number of topics:	10068
Total number of notes:	35879

9935.0. "nfs perormance - trucluster" by RTOMS::PARETIJ () Fri May 23 1997 14:58

VW are testing a Truecluster with engineering applications like MSC/NASTRAN.
their performance report follows 

"We have a memory channel cluster consisting of 4 x 4100 (each with 4
processors). The cluster has one common cluster disk (HSZ40, Stripeset with
2 x RZ29).

I observed a very poor performance with MSC/NASTRAN due to the I/O
characteristic of this application (NASTRAN produces a lot of I/O!). All the
NASTRAN scratch files where on the common cluster disk, which was NFS mounted on
each machine.

So I did some copy tests to find out how fast this cluster disk is. For that
purpose I used a 100 [MB] file and copied it between the local system disks and
the common cluster disk.

Here are the numbers, which I cannot explain and which are really bad:

1) Copy from system disk to system disk:
   19 sec
   
2) Copy from system disk to cluster disk (all machines, except the one which
   serves the cluster disk):
   130 sec
   
3) Copy from cluster disk to system disk (all machines, except the one which
   serves the cluster disk):
   120 sec for the first transfer, 19 sec for the following tests
   
4) Copy from system disk to cluster disk (for the machine serving the cluster
   disk):
   19 - 50 sec (wide spread!)
   
5) Copy from cluster disk to system disk (for the machine serving the cluster
   disk):
   22 sec
   
By the way: NFS copy time to a different machine is about the same as 2) and 3).

Is this behavior really normal or should anything be misconfiguered? "

T.R Title User Personal
Name Date Lines

9935.1 KITCHE::schott Eric R. Schott USG Product Management Fri May 23 1997 19:02 12

T.R	Title	User	Personal Name	Date	Lines
9935.1		KITCHE::schott	Eric R. Schott USG Product Management	`Fri May 23 1997 19:02`	12
	If you are using NFS loopbacks, you get the run thru all the NFS code.... You could consider using a disk service rather than an NFS service to improve this. Future cluster releases will have a cluster file system, which should do better at this. I suggest you move this to the smurf::ase notes conference...as cluster folks hang out there.

If you are using NFS loopbacks, you get the run thru all the
NFS code....

You could consider using a disk service rather than an NFS
service to improve this.

Future cluster releases will have a cluster file system, which should
do better at this.

I suggest you move this to the smurf::ase notes conference...as
cluster folks hang out there.