another quick joyous encounter: one of the co-workers was half-heartedly
beating his head for a month against a heavily-threaded java-based app
from a third-party vendor. the app ran on solaris sparc 8 and with 4-way
box it drove the sysload above 100 (!), while cpu utilization would
remain below 20% and prstat
showed more than a thousand threads.
in passing i asked him if he tried alternative thread library (another reference); i have always used it for our java app servers on solaris 8, but never saw any notable improvement, since there would be less than a hundred threads per JVM. in this case, however, the library solved the problem - the system load instantly dropped down to 1-2.
initially i liked M-to-M solaris 8 thread library, since its complexity was quite sexiful to anyone studying it theoretically, but apparently a simpler threading model is much more effective in the long run from many perspectives, and this is why it became the default in solaris 9.