Hi Bruce, I'm using NFS version 3 (nfs-utils-1.0.9) and both UDP and TCP protocols, with similar results. The entry in the exports file is: /data20 <network1>/24(rw,sync,no_root_squash) <network2>/24(rw,sync,no_root_squash) and in the client the mount is nfsserver:/data20 /data20 nfs rw,intr,hard or nfsserver:/data20 /data20 nfs rw,intr,hard,rsize=xyz,wsize=xyz where xyz could be any number multiple of 2 raging from 8192 to 32768. Thanks, Igor On Fri, Mar 6, 2009 at 12:42 PM, J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote: > On Thu, Mar 05, 2009 at 10:18:55AM -0700, Igor Sola wrote: >> Apologies in advance if this email is inappropriate for this mailing >> list. If so please let me know of a more suitable one. >> >> I noticed a significant performance difference when comparing a >> Centos5 (kernel 2.6.18) or Ubuntu (2.6.27-11) with a RH8 (kernel 2.4). >> All use NFS v3. >> The RH8 nfsd is faster than the 2.6 one even though it runs in an >> older, lower spec machine. >> >> See test details below >> >> Centos 5 x86_64 kernel 2.6.18-92.1.22.el5 >> >> RH8 2.4.20-30.8.multi_lun.smp #1 SMP i386 GNU/Linux >> >> 1Gb connection between machines >> >> File sizes are 1GB. The read speed of these files in the local machine >> filesystem (server) is around 98MBs in the Centos5 machine and around >> 65MBs in the old RH8 machine. >> >> .- First of all I wanted to measure how the network alone performs >> while using NFS. >> So in the server side I run a "cat" command on the 1GB file to >> /dev/null. At this point the file system has the 1GB file cached in >> memory. In the client side a "cat" on the same file gives me a speed >> of about 113MBs. Which is consistent with the 1Gb network. There is no >> performance hit. So far so good. >> >> >> .- The second test is reading from disk an uncache (in server and >> client) file. In the server and client I made sure I flushed the 1GB >> file from the memory. Between Centos5 machines the performance was >> about 35MBs (64% drop from the 98MBs disk read). Between RH8 machines >> the performance was 55MBs (15% drop from the 65MBs) >> >> This second test was repeated for ext2, ext3, xfs with no significant >> differences. >> >> Please note that the %iowait more than doubles when reading the file >> from the NFS partition vs the file system partition in the 2.6 kernels >> >> .- I run a third test just to make sure the problem is somehow related >> to the nfs daemon. I mounted the nfs partition local to the nfs server >> so the network was out the equation. I got the same poor results, no >> difference. >> >> Any ideas are most welcome, > > What version of nfs are you using, over what transport (tcp or udp)? > What are the export options? > > I wonder if something's going wrong with readahead. > > --b. > -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
