[rds-devel] [PATCH 03/21] RDS: Congestion-handling code

[Andrew Grover]

On Tue, Jan 27, 2009 at 5:10 AM, Evgeniy Polyakov <zbr at ioremap.net> wrote:
> On Mon, Jan 26, 2009 at 06:17:40PM -0800, Andy Grover (andy.grover at oracle.com) wrote:
>> +/*
>> + * Yes, a global lock.  It's used so infrequently that it's worth keeping it
>> + * global to simplify the locking.  It's only used in the following
>> + * circumstances:
>> + *
>> + *  - on connection buildup to associate a conn with its maps
>
> Is this a rare condition? Is this protocol only intended for the
> long-living connections and is not suitable for the cases when lots of
> them are created and teared down quickly?

Connections are long-lived. Imagine a cluster. RDS multiplexes all
sockets' datagrams between 2 hosts over a single transport-layer
connection, so if a node sends ONE datagram to another, an IB
connection is set up and sticks around indefinitely.

Regards -- Andy