At 08:28 PM 1/9/2006, Brian D. Ropers-Huilman wrote:
Simple question, likely a complicated answer. Cluster systems should have up
to three distinct file systems:

1) a place to compile and keep "stuff"
2) a bit of disk to use while code is running
3) potential a bit of disk with special characteristics, likely high-performance

Typically #1 would be called /home. Here at LSU, #2 is typically called
/var/scratch or /var/local/scratch. Then there is #3. Sometimes we call this
/scratch and sometimes /work.

My question: what do other folks call these bits of disk?

Our /home is the place for #1. If users/admins have specific programs that they want to make publically available, it can be published under admin-supervision in /usr/local. Given the number of nodes we have relative to network performance, and the potential aggregate bandwidth problem to the central server, it is not intended for any scratch file storage for running jobs, only for "real" data.

For small, non-bandwidth-sensitive local scratch (#2), we just use /tmp, local on each node. Small means there is at least 4GB available (up to 30GB on some nodes), but on a regular local disk.

Nodes designed for IO-intensive calculations have a /tmpfast filesystem where a raid-0 array is mounted (#3). Sizes are large (as large as needed for our current computational problem), throughput is high (such that IO/CPU bottlenecks are reasonably balanced for our typical problem sizes). If these nodes are used for non-IO-intensive jobs, it is nonetheless recommended to move regular scratch files (usually /tmp) to /tmpfast. A simple scripted check for the existance of this filesystem is sufficient and easy to implement to achieve this.

Our calculational needs do not include the permanent storage of very large datafiles, only large scratch files that are discarded after job completion.

Luc Vereecken





Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm

_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit 
http://www.beowulf.org/mailman/listinfo/beowulf

Reply via email to