B-heap performance and stability in practice

Anders Nordby anders at fupp.net
Tue Jun 15 12:07:51 CEST 2010


I had a discussion with PHK regarding the performance and stability of
r4646, aka the introduction of the B-heap. What I noticed after
upgrading from Varnish trunk version 4516 to 4755 which includes this

- CPU usage/load peaks went up generating warnings in Nagios.

- I get occasional hangs on HTTP, from 3 to 10 seconds or more, also
generating alerts.

- Varnish has space for more objects, but the total number of objects
fluctuates more.

- Disk wait time peaks are up by a couple of milliseconds (FYI I use
swap (malloc) backing on SSD in FreeBSD with max size set to 60 GB).

- Number of VM faults seems identical.

To back up my claims, I made a page with a graph comparison
between cache11 (which ran 4755 until june 8th when I downgraded to 4516
again) and cache12 (which has been running 4516 all the time):

My conclusion so far after downgrading from 4755 to 4516 is that in my
setup, Varnish trunk 4516 runs with less hiccups and load than 4755
which has the B-heap change. I wonder if anyone else has been testing
and measuring this change with setups that has a large number of
objects/large dataset. Any feedback welcome!



More information about the varnish-misc mailing list