Performance Tuning Ultrix 4.0

[Jon Giltner]

This in in regard to the article posted several weeks ago by Corey Satten
and Laurence Lundblade at the University of Washington about Tuning Ultrix
4.0 to encourage paging (over swapping).

Everything they described makes sense to me EXCEPT why Ultrix 4.0 seemed
so much slower than Ultrix 3.1 in the first place (the lotsfree, desfree, 
and minfree values did not change). What HAS changed is they way swap space
is allocated. Ultrix 4.0 has abandoned the familiar (4BSD) scheme of dmmin
and dmmax (where dmmin space is allocated sequentially, then if more is
needed dmmin*2 space is allocated sequentially and so on up to allocating
dmmax chunks sequentially). Instead Ultrix 4.0 has a configuration option
called swapfrag. Swap space is allocated swapfrag 512-byte blocks at a time.
By default this number is 64, which is 2*(default dmmin) under Ultrix 3.1.
Under the dmmin, dmmax scheme allocating about 1MB of swap space took
6 or 7 requests and the result was a few large chunks. Under the new scheme,
it appears that a 1MB allocation requires about 30 requests resulting in
30 32k chunks spread throughout the swap area. This could result in the
"thrashing" that some of us have experienced under 4.0.

It seems that we are now forced to strike a balance between swap space wastage
(if swapfrag is large and some processes are small) and  performance. Is
there something that I am missing? I'd like to know the motivation behind
this change.

Jon Giltner
University of Colorado
Computing and Network Services
giltner at pogo.Colorado.EDU
:wq