Hi All,
I have a question about the caching policy Lustre uses on
the server side. I''m using 1.4.7 and my current setting
is 1 MDS, 3 OST and 2 clients. My experience is that if
client 1 first fetches a file "foo" from Lustre to its
own /tmp directory and client 2 tries to fetch "foo" later,
client 2 will get the file faster. I wonder if Lustre is
using caching on the server side. I mean, data server, not
MDS. I searched this mail list and did not get clear clue.
Can anybody tell me if a Lustre server will save accessed
data in its memory or not? It is quite important in the
performance analysis of our research.
Thanks,
Zhe Zhang
____________________________________________________________________________________
Looking for a deal? Find great prices on flights and hotels with Yahoo!
FareChase.
http://farechase.yahoo.com/
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://mail.clusterfs.com/pipermail/lustre-devel/attachments/20070703/2876927a/attachment.html
On Jul 03, 2007 06:41 -0700, Zhe Zhang wrote:> I have a question about the caching policy Lustre uses on > the server side. I''m using 1.4.7 and my current setting > is 1 MDS, 3 OST and 2 clients. My experience is that if > client 1 first fetches a file "foo" from Lustre to its > own /tmp directory and client 2 tries to fetch "foo" later, > client 2 will get the file faster. I wonder if Lustre is > using caching on the server side. I mean, data server, not > MDS. I searched this mail list and did not get clear clue. > Can anybody tell me if a Lustre server will save accessed > data in its memory or not? It is quite important in the > performance analysis of our research.Lustre does NO data caching on the server. However, it DOES do caching of the file metadata, so this might speed the second access to the file. Cheers, Andreas -- Andreas Dilger Principal Software Engineer Cluster File Systems, Inc.
Hi Adrea, Thanks for the reply. What about read-ahead then? Is it the same as caching, and is performed on the client side only? Thanks, Zhe Zhang ----- Original Message ---- From: Andreas Dilger <adilger@clusterfs.com> To: Zhe Zhang <brucezhang_tech@yahoo.com> Cc: lustre-devel@clusterfs.com Sent: Tuesday, July 3, 2007 5:54:49 PM Subject: Re: [Lustre-devel] Lustre Server Caching On Jul 03, 2007 06:41 -0700, Zhe Zhang wrote:> I have a question about the caching policy Lustre uses on > the server side. I''m using 1.4.7 and my current setting > is 1 MDS, 3 OST and 2 clients. My experience is that if > client 1 first fetches a file "foo" from Lustre to its > own /tmp directory and client 2 tries to fetch "foo" later, > client 2 will get the file faster. I wonder if Lustre is > using caching on the server side. I mean, data server, not > MDS. I searched this mail list and did not get clear clue. > Can anybody tell me if a Lustre server will save accessed > data in its memory or not? It is quite important in the > performance analysis of our research.Lustre does NO data caching on the server. However, it DOES do caching of the file metadata, so this might speed the second access to the file. Cheers, Andreas -- Andreas Dilger Principal Software Engineer Cluster File Systems, Inc. ____________________________________________________________________________________ Don''t pick lemons. See all the new 2007 cars at Yahoo! Autos. http://autos.yahoo.com/new_cars.html -------------- next part -------------- An HTML attachment was scrubbed... URL: http://mail.clusterfs.com/pipermail/lustre-devel/attachments/20070709/1721af9c/attachment.html
On Jul 09, 2007 07:08 -0700, Zhe Zhang wrote:> Thanks for the reply. What about read-ahead then? Is it the same > as caching, and is performed on the client side only?The readahead is only on the client side, since most uses of Lustre have such a high load on the server that doing readahead on the server is a net loss of performance. The amount of client readahead is tuned on the client via /proc/fs/lustre/llite/*/max_read_ahead_mb. There is a second tunable /proc/fs/lustre/llite/*/max_read_ahead_whole_mb, which is the size of a file (in MB, or fraction thereof) below which the whole file will be read into the client cache regardless of the IO pattern.> ----- Original Message ---- > From: Andreas Dilger <adilger@clusterfs.com> > To: Zhe Zhang <brucezhang_tech@yahoo.com> > Cc: lustre-devel@clusterfs.com > Sent: Tuesday, July 3, 2007 5:54:49 PM > Subject: Re: [Lustre-devel] Lustre Server Caching > > On Jul 03, 2007 06:41 -0700, Zhe Zhang wrote: > > I have a question about the caching policy Lustre uses on > > the server side. I''m using 1.4.7 and my current setting > > is 1 MDS, 3 OST and 2 clients. My experience is that if > > client 1 first fetches a file "foo" from Lustre to its > > own /tmp directory and client 2 tries to fetch "foo" later, > > client 2 will get the file faster. I wonder if Lustre is > > using caching on the server side. I mean, data server, not > > MDS. I searched this mail list and did not get clear clue. > > Can anybody tell me if a Lustre server will save accessed > > data in its memory or not? It is quite important in the > > performance analysis of our research. > > Lustre does NO data caching on the server. However, it DOES do caching > of the file metadata, so this might speed the second access to the file. > > Cheers, Andreas > -- > Andreas Dilger > Principal Software Engineer > Cluster File Systems, Inc.Cheers, Andreas -- Andreas Dilger Principal Software Engineer Cluster File Systems, Inc.