[p4] has anyone used a RAM-SAN?

[Dan Bloch]

Me again.  I posted a brief and fairly neutral report on our RAM-SAN trial
yesterday, which I want to expand on and partially retract.  We did the
trial last week and we're still in the process of doing data analysis on it.

The trial was really quite positive, but (and this is the part which I've
come to realize between yesterday and today) it's hard to get metrics that
show it.  In particular, the numbers I was looking at were the times it took
to run the commands we're concerned with (principally large submits) and as
I reported they did improve but not dramatically.  However, what we're
really concerned with is the overall experience of the thousands of other
users on the server at the same time, and their experience is a function of
how long the big commands are holding database locks.  That time has gone
down, resulting in improvements which are hard to quantify. We're trying to
analyze this based on p4d process counts now.

Regarding David's analysis of RAM disks in general, it's correct but only up
to a certain point.  As he says, a RAM based disk will speed up the final
writes to the database (we didn't put the depot on the RAM-SAN), and this is
exactly what our big submits are bound by.  The other blocking commands we
see are big integrates, and these do a lot on non-contiguous reads, so again
they're helped by the RAM disk.

I'll post again if I learn anything else.

Regards,
Dan