Doug, First thanks for this. Some comments below.
On Wed, 29 Feb 2012 00:01:16 -0500 Doug Ledford <[email protected]> wrote: > There are two things that stand in the way of opensm being run on > redundant fabrics easily: > > 1) The opensm init script only starts one instance of opensm and opensm > will only work on one fabric per instance > 2) Even if you start multiple instances, you have to hand modify config > files for each instance and then when you upgrade the opensm rpm you > either loose your modifications or loose getting new default settings > > I worked around both of these issues, I've attached the files I used to > do so. > > First, I have an opensm init script that allows starting multiple opensm > instances. It supports configuring this in one of two ways: > > 1) Create multiple opensm.conf files, each with a numbered suffix (so > opensm.conf.1, opensm.conf.2, etc.) and it will start one opensm > instance per config file. This allows an admin to copy the default > config over and edit the things they need, and on rpm upgrade there will > be a new default opensm.conf file so they can diff between their edited > version and the new default and see if there are changes they need to > bring back in. This also allows for complete flexibility in setting up > the different fabrics, for instance you could use one type of routing on > one and a totally different type on the others. > > 2) Edit the file /etc/sysconfig/opensm and define more than one GUID in > the GUIDs variable. This will cause the opensm init script to > automatically start one instance per GUID, passing the GUID in on the > command line. I know you are going for ease of use here, which is good, however, I worry about this file becoming a redefinition of opensm.conf. > > For the most part, this works well. However, openmpi in particular > doesn't like you to have physically separate fabrics that have the same > subnet_prefix, and you can't specify a subnet_prefix on the command line > to go along with the GUIDs. So I wrote a patch for that and made the > init script unilaterally increment the subnet prefix for each different > GUID it's attaching to. If you only allow option 1 above this takes care of itself by making the admin configure his subnet prefixes in each config file as appropriate. The only down side is the loss of new configuration options as you upgrade. However, that is probably better taken care of by a default config file with each package. I mentioned this to Sasha years back and was denied since "you can always generate a new one with '-c'". :-( Alex would a default config file be acceptable? It would mean more work on your part. Ira > > All in all, we use the attached opensm file in /etc/sysconfig as the > standard place you put options belonging to an init script, we have the > opensm init script, the subnet_prefix patch I wrote, and with those > combined things work quite well. > > However, I will note that our init script does not (and will not ever) > play the passwordless root ssh stuff that upstream does. This is > considered a serious security risk on side. The idea that a customer > (let's say a wall street bank) should set up passwordless root ssh on > their cluster that's a backend to their web farm? Oh hell no... > > I might recommend that it is long since past time for that particular > misfeature of the upstream opensm init script to be done away with. > Personally, I would simply recommend that on failover from a primary to > a backup that it simply scan the fabric and build a "current guid2lid" > map from what it finds, then start updating from there. Or something > like that. But passwordless ssh...bleh. > > Oh, and while I've got your ear...is there a good reason the opensm libs > have been soname bumping so frequently? Is it not possible to extend > the APIs without soname bumps quite so often? > > -- > Doug Ledford <[email protected]> > GPG KeyID: 0E572FDD > http://people.redhat.com/dledford > -- Ira Weiny Member of Technical Staff Lawrence Livermore National Lab 925-423-8008 [email protected] -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
