From: =?iso-8859-1?Q?Brasseur_Val=E9ry?= Subject: RE: nfs unlink perf ! Date: Tue, 13 Aug 2002 18:52:57 +0200 Sender: nfs-admin@lists.sourceforge.net Message-ID: <8D7D56C6ED3DD411998D009027DC685E0874E226@srv-grp-s1.segin.com> Mime-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Cc: nfs@lists.sourceforge.net Return-path: Received: from hermes8.atos-group.com ([160.92.18.57] helo=hermes8.segin.com) by usw-sf-list1.sourceforge.net with esmtp (Exim 3.31-VA-mm2 #1 (Debian)) id 17eeuZ-00033G-00 for ; Tue, 13 Aug 2002 09:52:51 -0700 To: "'Lever, Charles'" Errors-To: nfs-admin@lists.sourceforge.net List-Help: List-Post: List-Subscribe: , List-Id: Discussion of NFS under Linux development, interoperability, and testing. List-Unsubscribe: , List-Archive: > -----Original Message----- > From: Lever, Charles [mailto:Charles.Lever@netapp.com] > Sent: Tuesday, August 13, 2002 6:21 PM > To: 'Brasseur Val=E9ry' > Cc: nfs@lists.sourceforge.net > Subject: RE: [NFS] nfs unlink perf ! >=20 >=20 > hi brasseur- >=20 > > I check nfsstat, and see a bunch of retrans... > > but now when then following your suggestion I 've try to put : > > tcp + rsize/wsize=3D32768 and : > > * take the same time > > * hang on D state under moderate load so I need to reboot=20 > > very frequently. >=20 > if "nfsstat -c" shows lots of retransmissions and you are using > NFS over UDP, follow these instructions to increase the transport > socket buffer size: >=20 > 1. Become root on your client > 2. cd into /proc/sys/net/core > 3. echo 262143 > rmem_max > 4. echo 262143 > wmem_max > 5. echo 262143 > rmem_default > 6. echo 262143 > wmem_default > 7. Remount your NFS file systems on the client >=20 already done that ! it's my default config ! > if that doesn't clear your problem, then you probably have a faulty > network -- swap your cabling, look for NIC errors on your client > and filer, check the statistics on your switch, and be sure duplex > and autonegotiation are correctly set. you can ask your NetApp SE > for help diagnosing network problems on the filer if you need it. no errors are reported by NICs nor switch ! full-duplex is force=20 on all devices (client,filer,switches ...) >=20 > also try testing from other clients on other networks, if you can. >=20 > > my previous setup was : > > * nfs v3, rsize/wsize=3D8192, udp > >=20 > > my client (DL360) and server (F880) are connected through=20 > a dedicated > > switch each with 100FD network link to the switch. > >=20 > > any ideas ? >=20 > if the above suggestions don't help, get a packet trace. you can use > "tcpdump" on your client, or you can try "pktt" on the filer. > ("pktt list" on the filer's console should tell you what you need to > know). pktt produces a trace file in tcpdump format, so you can use > tcpdump on your clients to look at the dump. >=20 > send me the dump if you'd like help. >=20 > > > -----Original Message----- > > > From: Lever, Charles [mailto:Charles.Lever@netapp.com] > > > Sent: Monday, August 12, 2002 5:48 PM > > > To: 'Brasseur Val=E9ry'; nfs@lists.sourceforge.net > > > Subject: RE: [NFS] nfs unlink perf ! > > >=20 > > >=20 > > > link RPCs can be very large (much larger than wsize, for=20 > instance). > > > if you use UDP and your network drops lots of packets, large link > > > requests will probably require several retransmissions to=20 > complete. > > >=20 > > > enlarge your transport socket buffer size on the client (see FAQ) > > > and/or try NFS over TCP. also examine your network to see if it > > > is clean, and is working optimally. if your client is attached > > > via a more slow link than your server, that can be the source of > > > bursty packet loss. > > >=20 > > > > -----Original Message----- > > > > From: Brasseur Val=E9ry [mailto:Valery.Brasseur@atosorigin.com] > > > > Sent: Monday, August 12, 2002 8:09 AM > > > > To: nfs@lists.sourceforge.net > > > > Subject: [NFS] nfs unlink perf ! > > > >=20 > > > >=20 > > > > I am testing an application which to a lot of link/unlink > > > >=20 > > > > server is a NetApp filer, client is linux 2.4.19+nfs patches. > > > >=20 > > > > under load it seems that the unlink call is very long when=20 > > > > all over calls > > > > response time seems to be OK. > > > > is there any problems with the unlink code ? > > > >=20 > > > > thanks > > > >=20 > > > >=20 > > > >=20 > > > > ------------------------------------------------------- > > > > This sf.net email is sponsored by:ThinkGeek > > > > Welcome to geek heaven. > > > > http://thinkgeek.com/sf > > > > _______________________________________________ > > > > NFS maillist - NFS@lists.sourceforge.net > > > > https://lists.sourceforge.net/lists/listinfo/nfs > > > >=20 > > >=20 > >=20 >=20 ------------------------------------------------------- This sf.net email is sponsored by: Dice - The leading online job board for high-tech professionals. Search and apply for tech jobs today! http://seeker.dice.com/seeker.epl?rel_code=31 _______________________________________________ NFS maillist - NFS@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/nfs