Agreed. There has been a ton of work in DSM efforts, there are many
PhD thesis's to read on this topic, and they all conclude that the
performance is horrible on any application which has any significant
amount of shared data (whether that data is shared data or shared locks
makes no difference).
I've personally sat through one of these thesis's and painfully watched
this otherwise extremely intelligent person beat this idea to death
until there was no possible hope left (he got his thesis, by the way,
from Stanford). He was doing this in IRIX, on SGI machines using full
duplex HIPPI (100MB/sec, any packet size you want, ping pong latency of
about 300 usecs).
The problem is that people program their systems assuming that locks
will be fast, on the order of a microsecond in the worst case remote
cache fetch. If it takes 10,000 microseconds to fetch a lock using
DSM, and that number is not unreasonable, you can easily see there will
be problems.
DSM sucks. It really does. It's one of those things that you want to
believe will work because it will make your life easy and then it doesn't
in practice. So you waste all this time to figure that out and end up
having to rework your app in terms of messages anyway.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/