Jan Wieck <JanWieck@Yahoo.com> writes:
>> That's irrelevant to the problem, though. Unless the ARC code uses data
>> structures that are more amenable to localized locking than the old
>> global buffer freelist. (Jan?)
> the strategy itself does no locking at all. Like the old LRU code it
> simply assumes that the buffer manager holds the lock during calls.
Okay, I suspected as much but wasn't sure.
Manfred's numbers definitely say that we need to find a way to break
down the BufMgrLock into multiple finer-grain locks. We already have
all those per-buffer LWLocks, but I don't see how to apply those to
the problem of managing the global lookup and replacement datastructures.
Anyone see an attack path here?
regards, tom lane