Re: spinlock contention

On 23.06.2011 18:42, Robert Haas wrote:
> ProcArrayLock looks like a tougher nut to crack - there's simply no
> way, with the system we have right now, that you can take a snapshot
> without locking the list of running processes. I'm not sure what to
> do about that, but we're probably going to have to come up with
> something, because it seems clear that once we eliminate the lock
> manager LWLock contention, this is a major bottleneck.

ProcArrayLock is currently held for a relatively long period of time
when a snapshot is taken, especially if there's a lot of backends. There
are three operations to the procarray:

1. Advertising a new xid belonging to my backend.
2. Ending a transaction.
3. Getting a snapshot.

Advertising a new xid is currently done without a lock, assuming that
setting an xid in shared memory is atomic. To end a transaction, you
acquire ProcArrayLock in exclusive mode. To get a snapshot, you acquire
ProcArrayLock in shared mode, and scan the whole procarray.

I wonder if it would be a better tradeoff to keep a "materialized"
snapshot in shared memory that's updated every time a new xid is
assigned or a transaction ends. Getting a snapshot would become much
cheaper, as you could just memcpy the ready-built snapshot from shared
memory into local memory.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com