Latest news: I've bodged the installation, but I'm having problems with nfs.mount now.
Since everything worked with nfs-rdma on the development server, I manually copied the kerel-ib rpm file from there, and installed that on top of the default infiniband install on the live server, using the command: # rpm -iv --replacepkgs kernel-ib....... With that done modprobe xprtrdma now works. Since the kernel-ib file appears to be the only one modified as nfsrdma support is added, I assume this is all that's needed. Can anybody confirm if this is ok, or whether I'm going to break things horribly by doing this? I also have nfs v1.1.6 installed, but it doesn't seem to understand the rdma option. The system is still using tcp by default, and if I try to use rdma, I get this: # mount -o rdma, port=20049 server:path /mnt mount.nfs: Unsupported nfs mount option: rdma However, nfs is definitely the correct version: # mount.nfs -V mount.nfs (linux nfs-utils 1.1.6) Even using the mount.rnfs utility supplied with OFED doesn't help: # rpm -iv rnfs-utils-1.1.5-2.OFED.i386.rpm # mount.rnfs server:path /mnt -o rdma,port=20049 mount.rnfs: Unsupported nfs mount option: rdma And if I use mount.rnfs without the rdma option, "cat /proc/mounts" reports that it's just using tcp. I'm getting read speeds of around 400MB/s using regular NFS which isn't bad, but I know I can achieve over 900MB/s with this server over Infiniband, so there's still some way to go. Ross On Fri, Jun 26, 2009 at 2:06 PM, Ross Smith<[email protected]> wrote: > Ok, I've made a little more progress. > > The issue with nfs-utils was solved by installing efsprogs-devel, and > mount.nfs is now reporting that it's running v1.1.6. > > I still can't install OFED with nfs-rdma on the live server though, it > appears that it's unable to use the rpm package to install with nfs > support. > > I've now discovered how to use the ofed.conf file, and this is the > procedure I'm following: > > - On the development server, I run install.pl and perform a standard install > - Once that finishes, I edit the resulting ofed.conf and add nfsrdma=y > - Still on the development server I run: > # install.pl -c ofed.conf > - That can be seen to install most items from the rpm, and to > recompile kernel-ib with the nfsrdma option. > - I then package up the OFED folder with tar, and send it to the live server > - On the live server I run the same command: > # install.pl -c ofed.conf > > Despite it running with exactly the same options, I get the error: > > "/lib/modules/2.6.18......./build/scripts is required to build kernel-ib RPM. > Please install the corresponding kernel-source or kerenel-devel RPM." > > If I exclude the nfsrdma option and repeat all the above steps, the > script installs fine from the pre-compiled RPM. > > It looks to me like a problem with the install script when choosing nfs-rdma. > > Does anybody have any suggestions as to how I can work around this? > Am I able to simply install the RPM files manually? > > Ross > > > > On Fri, Jun 26, 2009 at 12:25 PM, Ross Smith<[email protected]> wrote: >> Thanks Robert, >> >> Looking through your notes, we've followed similar steps. I should >> mention that I do have Infiniband working on this server, including >> ipoib. It's purely the nfs-rdma part I'm struggling with. >> >> My latest attempt has been to attempt a custom install, selecting >> nfs-rdma manually. That worked fine on the development server, but >> when I attempted to run it on the live server I had an error: >> >> "/lib/modules/....../build/scripts is required to build kernel-ib RPM. >> Please install the corresponding kernel-source or kernel-devel RPM. >> tk rpm is required to install ibutils" >> >> The problem is, for XenServer, you can't get the kernel-source on the >> live server, instead Citrix ship a DDK (driver development kit) >> virtual machine, which is a pre-built system with all the kernel >> sources you need. >> >> OFED builds fine on the DDK machine, and for the default installation, >> I can compile everything there and then transfer the OFED folder (with >> the compiled RPM files) to the live server. At that point install.pl >> runs fine, without needing to re-compile things. >> >> However, it appears that when I choose some elements in the custom >> install, it is not using these RPM files, and is trying to compile >> kernel-ib again from scratch. >> >> I have attached a file listing the custom modules I have chosen in >> case it helps. >> >> Ross >> >> >> >> On Fri, Jun 26, 2009 at 11:42 AM, Robert Dunkley<[email protected]> wrote: >>> Hi Ross, >>> >>> I installed OFED 1.3.1 (Older I know) with other options OK on a Centos >>> 5.2 Dom0 (Separate Xen install on Centos, not "Xen Server"), here is >>> some notes I took (I used the GUI installer) >>> >>> Hope this helps, >>> >>> Rob >>> >>> 1. Download the OFED package from >>> http://www.openfabrics.org/downloads/OFED/ >>> 2. Extract it: tar -xvzf OFED-1.3.1.tgz >>> 3. Check for the dependencies by running ./install.pl and choosing >>> documentation ("0"). >>> 4. Install all required dependencies using yum install package nam. >>> For Centos dependencies try: yum install gcc libstdc++-devel >>> libsysfs-devel tcl tcl-devel tk pciutils-devel kernel-devel rpm-build >>> gcc libtool bison flex tcl-devel swig gcc-c++ libtool kernel-xen-devel >>> zlib-devel You may need to run the yum command three times (It takes a >>> few runs for all the dependencies to sort themselves out) >>> 5. It is recommended you now upgrade the kernel to make sure the >>> sources match it. For centos do: yum upgrade kernel (Or yum upgrade >>> kernel-xen for Xen setups) >>> 6. Run ./install.pl again but this time choose to install ("2") and >>> choose custom install ("4"). See the appendix for a recommendation on >>> which modules to install. >>> 7. Install should now complete, if it fails check the log (Most >>> likely you have not installed a required software package - go back to >>> step 3) >>> 8. Say yes to configuring IPOIB (If you plan to use it) and >>> manually set the IP and subnet. When it asks for a network it wants the >>> base network address (If given an IP of 192.168.10.12 on a 255.255.255.0 >>> subnet then the "Network" would be 192.168.10.0) >>> 9. Reboot >>> 10. Check the install using /etc/init.d/openibd status. This should >>> show one interface for each Infiniband port (First will be "ib0"). You >>> should also see quite a few modules loaded, the important ones are >>> (mlx4_ib & mlx4_core - QLogic drivers, mthca - Mellanox driver, ib_core >>> & ib_addr - core services, rdma_ucm & rdma_cm - RDMA, ib_cm & ib_ipoib - >>> IPOIB and ib_uverbs - Direct/Verb) >>> 11. Run ibstat - This should display the installed devices and their >>> firmware versions and hardware / GUI IDs. If you need to update the >>> firmware for Mellanox see >>> http://www.mellanox.com/support/firmware_download.php I find it easier >>> to update using Windows and the WinOF software (Windows equivalent of >>> OFED - Available here: http://www.openfabrics.org/downloads/WinOF/ ). >>> >>> -----Original Message----- >>> From: [email protected] >>> [mailto:[email protected]] On Behalf Of Ross Smith >>> Sent: 26 June 2009 11:36 >>> To: [email protected]; [email protected] >>> Cc: Yehonatan Yossef >>> Subject: [ofa-general] Installing nfs-rdma with OFED-1.4.1 >>> >>> Hello everyone, >>> >>> Can anybody help me get nfs-rdma installed on CentOS 5.2 (actually >>> Citrix XenServer), with OFED-1.4.1? >>> >>> The default install of OFED doesn't include nfs-rdma, and I'm >>> struggling to work out which components I do and don't need for a >>> custom build. >>> >>> What I would like to include is: >>> - Core infiniband components >>> - Drivers for Mellanox Infinihost III Lx (MHES14) >>> - IPoIB >>> - nfs-rdma >>> - troubleshooting and diagnostic tools (ibstat, ibtracert, ibswitches, >>> ibhosts, etc...) >>> >>> My last attempt was following the instructions from OFED 1.4 - I >>> edited ofed.conf to add nfsrdma=y, and attempted a default install. >>> However after running install.pl, that change appears to have been >>> overwritten. >>> >>> I am also having problems updating mount.nfs on this machine. Both >>> nfs-utils-1.1.6 and nfs-utils-1.2.0 fail to install with the error: >>> "blkid/blkid.h: No such file or directory", and I cannot yet find a >>> source for this file. >>> >>> Finally, my understanding is that nfs-rdma is in beta status for >>> CentOS 5.2, but I couldn't see any open bugs on the tracker, are there >>> any issues I should be aware of? >>> >>> thanks, >>> >>> Ross >>> _______________________________________________ >>> general mailing list >>> [email protected] >>> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general >>> >>> To unsubscribe, please visit >>> http://openib.org/mailman/listinfo/openib-general >>> >>> The SAQ Group >>> >>> Registered Office: 18 Chapel Street, Petersfield, Hampshire GU32 3DZ >>> SAQ is the trading name of SEMTEC Limited. Registered in England & Wales >>> Company Number: 06481952 >>> >>> http://www.saqnet.co.uk AS29219 >>> >>> SAQ Group Delivers high quality, honestly priced communication and I.T. >>> services to UK Business. >>> >>> Broadband : Domains : Email : Hosting : CoLo : Servers : Racks : Transit : >>> Backups : Managed Networks : Remote Support. >>> >>> ISPA Member >>> >>> >> > _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
