Hello,

I just put together a Beowulf network at home, using Slackware 12.2 as the server. Two other PCs boot from CDs and become diskless clients, so it's a total of three PCs. Everything works fine, so far so good.

My question is this. It looks like it's fairly likely I'll be asked to schlep my PC to a local school and demonstrate the Beowulf using their PCs. Now, I've heard somewhere that there's a default but easily changeable limit in Slackware somewhere on how many NFS clients a server will support, on the order of a half-dozen or so. Problem is, I can't remember where I heard that, and Googing doesn't seem to land me on any informative pages. I don't even remember if whoever or whatever I heard it from seemed particularly knowledgeable, just that I did hear it somewhere.

So, before I try to boot 25 PCs off this box and embarrass myself by running into some max_clients=6 limit ... is there such a limit? And if so, where in the source and/or config files would I go to change it? Thanks!

SAM

There is not a direct limit on the number of clients.

However, check this out:
http://nfs.sourceforge.net/#section_b

You may want to make sure that you are starting a few extra threads in order to be sure an avoid embarrassment . . . not from denial of service but slowness.

Also, how fast is your hard drive or RAID? I hope it is something with some speed and not something slow.

Really, the entire page at http://nfs.sourceforge.net/ is a good read. I had to dig into the slackware sources on the 13 install dvd in order to figure out what software package exactly the nfs daemons and utilities were.

That's good to hear; thanks!

I'll keep the extra threads trick handy, but don't think I'll need it. In this sort of configuration, the client nodes have no humans sitting in front of them working, no GUIs running, etc., and often don't even have keyboards or monitors. For all practical purposes, all the NFS traffic occurs during booting and execution of /etc/rc scripts. After that, they're running just enough of a stripped-down OS to support slave processes which will receive data from the master via TCP/IP, do a *lot* of computations on that data, and send back the results to the master. The only process doing file I/O is the master, which lives on the server where the hard drivers are located, so no problem there. So, each slave gets a small packet containing its assigned data to work on, does a lot of number crunching on that data, and some time later sends it back to the master which coordinates everything, turning the entire network into one big virtual parallel machine. I'm not worried at all about file I/O or network bandwidth bottlenecks (and I've already spec'd out things on my home network to confirm that there are no worries and that the VPM software itself will scale up); what I was concerned with was that the server might simply drop a client because of some hard-coded limit, and then sometime later the client might do a query just to re-verify its NFS connection, find no server, and freak out. Thanks again!

SAM