The last time I looked at the cost of swap to vdisk, at 1,000 per
second, used 10% of an 890 processor. It's very hard to constrain a
system to swap this much, this was in the lab pushing limits not
normally pushed. With z10 IFL significantly faster, swapping to vdisk
would not be a significant cost.
The largest performance problem facing us today is storage, as IBM
Software Group has decided to put polling back into their applications
(remember the hertz timer in Linux we eliminated in 2003? - it's back
courtesy of IBM applications). With polling, the over commit ratio you
can attain is now about 1.5 - so reducing Linux storage sizes and
causing some swap means more Linux servers per installed storage.
Robert J Brenneman wrote:
Just a guess till the experts chime in:
Linux disk I/O activity requires more CPU time than traditional Z
Operating systems - so when one guest starts driving 5000 I/O ops per
second to the swap device ( FBA mode vdisk in my case ) that in itself
consumes a big chunk of CPU. Then there's the additional time spent in
the linux kernel itself deciding what needs to go out to swap and what
needs to come back in.
let me re-emphasize this is a guess - I'd like to know the answer to this too.