filesystem activity causes 4.0 to bog down?

Rick Ace pixar!rta at ucbvax.berkeley.edu
Wed Feb 22 11:12:55 AEST 1989


Upon porting some of our software to SunOS 4.0 recently, we observed a
disturbing slowdown of the window system when running a certain
file-intensive application.  To get a better handle on the nature of this,
ahh, behavior, we experimented a little, playing with the window system
with other things happening concurrently.  All tests were made on SunOS
4.0 (not sure of the exact beta-ness of the release).

1.  Sun-4 (100-series) with a SCSI disk and 8Mbytes of memory.
No activity on system.  Window system is responsive (windows open and
close quickly).

2.  Same system.
Experimental load:  dd if=/dev/rsd0g of=/dev/null bs=10k
Window system slightly sluggish, but no severe degradation.

3.  Same system.
Experimental load:  CPU-bound program in tight loop, no I/O.
Window system again slightly sluggish, but not severe.

4.  Same system.
Experimental load:  cp largefile newfile
Now the window system bogs down real bad.  Typing ^Z in the
window where the "cp" was running had no effect!  Trying to
open or close other windows experienced extreme delays.

5.  Sun-4/280 with Hitachi disk, Xylogics ctlr, 16 Mbytes.
Experimental load:  cp largefile newfile
Same symptoms as test # 4.

The results suggested this hypothesis:  the new VM scheme in 4.0 uses main
physical memory as a combination file buffer cache and page storage for
"in-memory" processes.  The "cp" program reads blocks from the source file
and copies them to the destination file, demanding memory for buffer cache
purposes.  The kernel memory manager throws processes out of memory to
make room for the file blocks being copied by cp.  When the user tries to
use the window system, that causes processes to wake up and want to get
swapped back in, and the process swapper must now compete with file
management for physical memory, and somehow the file manager seems to be
making life very tough for the swapper.

How do y'all feel about this?  Have you noticed this type of behavior on
your systems?  Does the theory seem plausible?

Rick Ace
Pixar
3240 Kerner Blvd, San Rafael CA 94901
...!{sun,ucbvax}!pixar!rta



More information about the Comp.sys.sun mailing list