MMA, HP9000/7xx, multiple users
- To: mathgroup at yoda.physics.unc.edu
- Subject: MMA, HP9000/7xx, multiple users
- From: David Harrison <harrison%faraday at yoda.physics.unc.edu>
- Date: Thu, 21 Jan 1993 14:28:43 -0500
This is a very specific request for information involving MMA, HP9000/7xx computers, X-terminals, and multiple users. Apologies to those for whom it is too specific to be of interest or use. We are running 27 X-terminals off a thinwire ether served by an HP9000/750 running HPUX-8.07. We have many applications, but our most resource intensive is MMA 2.1. In a nutshell, when multiple MMA and especially motifps processes are running, response on our X-terminals slows to a crawl. (Crawl means up to 15 seconds to get a response to a carriage return in an xterm.) The problem effects all X-terminals on the network, not just the one running all those processes. Our ethernet sniffer indicates there are no problems with the network: 15% utilisation max, small number of collisons/packet, trivial number of runts/packet, no frame errors. The loadavg on the HP is reasonable, we have some memory left usually, response on serial terminals is fine: just the X-terminals are slow. The network is attached to our second ether controller, on the EISA bus. The motherboard ether is connected via an AUI to our campus backbone. We have HP working on the problem from their side. Nonetheless, the thought occurs that perhaps we are seeing something MMA specific. At the moment, for example, we have 150 user processes running, loadavg is 2.5, of the 150 user processes 6 are rather big xview applications, no MMA's running, and response is excellent. Substitute MMA and motifps for those xview applications and it response goes to a crawl, although the time to actually do a calculation seems unaffected. In particular, those motifps processes all take between 5 and 7 seconds of cpu (a very long time for a box of this horsepower) and consume between 1911 blocks (~= 1 megabyte) and 7644K of memory depending on whether you believe 'ps -l' or 'top'. When we were running the same applications with the same physical configuration with MMA 2.0 on a Sun 4/280S we didn't see this problem (although we sure saw some other ones!). Anybody out there with a sufficiently similar configuration who can give hints, fixes, etc. gets a big 'thanx'. -- David Harrison | "Why do three tones make a Dept. of Physics, Univ. of Toronto | triad and not a triangle?" Inet: harrison at faraday.physics.utoronto.ca | -- Mach