[Gluster-devel] memory leaks

Anand Babu ab at zresearch.com
Fri Mar 9 03:11:23 UTC 2007


,----
| Using stat-prefetch on the server seems easily fatal; just ls or du
| a subdirectory and the glusterfsd processes die instantly.  I'm not
| sure if stat-prefetch really makes sense on the server, anyway,
| though.  It seems to work fine on the client.
`----
stat-prefetch was designed to reduce network operations, which means
it accelerates client's sequential stat lookup calls. For example, if
you do "ls -l", after fetching dir-entries in bulk, it calls stat
operation for each and every file sequentially. If you have thousands
of files in a dir, you can imagine the number of network
operations. With stat-prefetch, when you issue a stat call for a file,
it returns stat info for all the files in the dir, no matter
what. Consecutive stat calls for the rest of the files in that dir
will use the client side cache. This stat-info is cached until
"cache-seconds" expires. 

Having said that, it is important to make sure translators work
seamlessly on both server and client. We never know, GlusterFS is
programmable and it leads to many possibilities. Can you please file a
bug report for stat-prefetch on server side.

,----
| Read-ahead does, of course, have the memory leak on server as well
| as client.
`----
Amar is working in it. This task has been assigned highest priority.

-- 
Anand Babu 
GPG Key ID: 0x62E15A31
Blog [http://ab.freeshell.org]              
The GNU Operating System [http://www.gnu.org]  






More information about the Gluster-devel mailing list