home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.unix.sys5.r4
- Path: sparky!uunet!spool.mu.edu!uwm.edu!rpi!newsserver.pixel.kodak.com!laidbak!tellab5!mcdchg!mcdphx!udc!avonruff
- From: avonruff@urbana.mcd.mot.com (Al Von Ruff)
- Subject: Re: Unix System Crashing!! Why?
- Message-ID: <1992Nov9.195145.8863@urbana.mcd.mot.com>
- Sender: news@urbana.mcd.mot.com (News)
- Nntp-Posting-Host: predator.urbana.mcd.mot.com
- Organization: Motorola Computer Group, Urbana Design Center
- References: <1992Nov3.130757.20385@cbfsb.cb.att.com> <1992Nov5.140301.20874@urbana.mcd.mot.com> <900@dsbc.icl.co.uk>
- Date: Mon, 9 Nov 1992 19:51:45 GMT
- Lines: 66
-
- In article <900@dsbc.icl.co.uk> mbm@dsbc.icl.co.uk (Malcolm Mladenovic) writes:
-
- >>1) strthresh is tuned too low for your system use. If your machine
- >> crashes at approximately the same time of day (i.e: during backups,
- >> while sending/receiving news, during high UUCP traffic, running heavy
- >> networking loads), then this is probably the case. strthresh is a
- >> kernel tunable which designates the maximum number of bytes which
- >> the STREAMS subsystem is allowed to use. This parameter has a
- >> notoriously low default setting (around 2 Megabytes). I would recommend
- >> setting strthresh to a value of 50 percent of physical memory.
- >
- >This seems an unnecessarily high a value for large configurations - the machine
- >that I'm posting from has strthresh set to 4MB. It is a dual processor sparc
- >running svr4 with 128MB of physical memory. The machine is regularly running
- >with large amounts of networking traffic but no STREAMS resources problems.
- >This ought to be sufficient unless you are running with many TCP connections
- >doing bulk data transfer at the same time. [It doesn't actually do any harm
- >to have a large value, if you do have a problem it will increase the length
- >of time the system stays up, though performance will degrade over time since
- >less and less memory will be available for the buffer cache and other uses.]
-
- I have found that the strthresh setting required for a particular installation
- depends greatly on the type of use encountered. The 50 percent setting is
- actually quite close to USL's suggested setting:
-
- "STRTHRESH should be set to about 1/4 to 1/2 of the total system
- memory. The default of 2000000 (approximately 2 megabytes) is a
- good maximum for a 4 megabyte system."
-
- The above is from the "Tunable Parameters" section of the SAG (chapter 10).
-
- In practice, I have found that 2 Meg is sufficient for single user systems,
- or systems with light multi-user loads. I have found that 4 Meg is sufficient
- for most other multi-user systems - but I have seen several notable exceptions.
-
- For instance, our main source configuration backbone (A quad 88K with 64 MB of
- physical memory) has a typical average load of 300 to 600 processes, about 80
- to 100 users, with file systems NFS mounted via T1 link to somewhere between
- 40 to 60 remote hosts. This system has an extremely high STREAMS usage profile,
- and 4 Meg of STREAMS buffers is simply inadequate.
-
- I have also observed a dual CPU system with 128 MB with four modems (lots of
- ldterms and STREAMS-based tty drivers) connected to uunet (serving as
- a UUCP hub) die of resource starvation unless strthresh was boosted well
- past 8 MB.
-
- As you stated, there isn't any penalty for configuring "too much" STREAMS
- resources (as long as you stay well below physical memory limits). The real
- drawback that I have seen is that it prolongs system death in cases where
- real problems exist, dragging out the debug time for that particular problem.
- As you also stated, these systems are usually pretty easy to identify due
- to their dog-slow performance.
-
- > .... If Strcount increases over time
- >with no corresponding change in the system workload you probably have a
- >memory leak in either the STREAMS subsystem itself or in one of the STREAMS
- >modules or drivers. If this is the case you need to report the problem
- >to your vendor. It may also aloow you to do planned, clean reloads at a more
- >convenient time than just when the system locks up.
-
- Agreed.
-
- -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
- Al von Ruff Motorola Microcomputer Division, Urbana Design Center
- Voice: 217.384.8553 1101 East University Avenue, Urbana, IL 61801 USA
- Internet: avonruff@urbana.mcd.mot.com UUCPnet: uiucuxc!udc!avonruff
-