Bill Sommerfeld wrote:

> I suspect that's what going on here is a scaling problem; systems are
> dirtying data at ever-faster memory speeds, and then flushing it at
> disk speeds only every 30s, leading to a significant pause..
> 
> It's not immediately clear what's causing the other processes to block
> during the sync; ffs_update is one likely place (for instance, if the
> bread() call it makes misses in the cache..).

Soft updates anyone?  Bueller?

When you had 15 second syncs, did the pauses shrink correspondingly?
If so, a tacky workaround would be to run "update 10" (or some random
value) rather than plain "update".

Simon.