Disk Striping - B-2000

[Charles Fear]

I have a B-2000 with 2 disks... will I see better performance form the Disk I/O if I stripe both disks... Then the I/O would be spread accross both disks for tha abaqus application data...

Would the down side to this be that the OS would also be striped across both disks?

[Steven E. Protter]

You would see better performance and reliability if you used mirror/ux to do a raid 1 copy.

I think you may actually slow peformance down by trying to stripe, but can not confirm that.

Your best bet is to run some performance tests, do the striping and then run the same tests.

Attaching some scritps that will help you collect performance data.

SEP

[James A. Donovan]

You would see better performance with striping. The downside would be that since you are not mirroring the disks, if one disk goes bad, then you have to rebuild the machine from backups.

[Charles Fear]

Ultimately what we are building is a cluster of 20 B-2000's running Abaqus. There will be a node 0 or master node that will have a striped disk set for the majority of the processing. Ths advantage of tyring to increase the performance of the other 19 nodes is to reduce the time for the -re and post processing.

Each node needs to have a 5 - 7 GB disk area that the post and pre processing happens on... With only two disks in the system.. I am looking for ways to make these filesystems as fast as possible.

[Jeff Schussele]

Hi Charles,

You know you are talking about 700 series workstation level systems here. They're not going to be barn burners any way you slice it.
BUT if you're *main* concern is speed, then I would look at dual SCSI HBAs & have each drive on a separate HBA. The key would be to keep the vg00 & vg_app VG traffic apart. Any way you stripe or mirror these two disks you have to keep the app & the OS together. This not only hampers througput but it's dangerous to the OS as a corruption in the app FS could hammer vg00 and panic the box.
You could put a tape or CD/DVD on the same bus as vg00, but NOT the app traffic. Keep it on it's own bus - I think that's the fastest you're going to get.

My 2 cents,
Jeff

[Tim D Fulford]

Hi

I used to work with ABAQUS a few years ago.. Personally (my opionion dofferes from some of the above) I would agree striping is better for IO performance than mirroring (or nothing)..

I did hear that SDRC-Ideas developed a method whereby ALL the Gausiean elimination was flushed to disk. This allowed large node models to be analysed on relatively small systems ... but VERY VERY VERY SLOWLY.. (typically memory is 1,000+ times faster than disk). If ABAQUS have done the same then you may be stuffed. In which case I'd stripe the "dump" logical volume, but at best you will halve analysis times.

That said, I very much doubt the disk is
the bottleneck. As you know ABAQUS is finite element program/package & mostly works by reading in a (relatively) small ASCII file and within memory, constructing matricies, inverting these matricies (Gausiean eleimination etc).. blah blah. The point I'm trying to make is that you will be doing the majority of the number crunching & matrix strorage work within memory. So if you are seeing high disk utilisation I suspect it would be the root disks being used to swap out because of lack of memory... if so a process called vhand would be running fairly constantly.

That said...
1 - Are you seeing swapping (vhand)
2 - Why do you want to speed up IO.. is it to reduce analysis times? Or something else.

Regards

Tim

[Charles Fear]

To answer your question... Yeah I know they are workstations... This is what the company has to work with... I have another company that is using Itanium RX2600 1.4 Ghz {2} for this same funciton.

I am not seeing anything as far as performance bottlenecks yet as they are just getting started ... just trying to prevent them as much as possible