llonergan@greenplum.com ("Luke Lonergan") writes:
> Christopher,
>
> On 3/23/06 6:22 PM, "Christopher Browne" <cbbrowne@acm.org> wrote:
>
>> Question: Does the Bizgress/MPP use threading for this concurrency?
>> Or forking?
>>
>> If it does so via forking, that's more portable, and less dependent on
>> specific complexities of threading implementations (which amounts to
>> non-portability ;-)).
>
> OK - I'll byte:
>
> It's process based, we fork backends at slice points in the execution plan.
By "slice points", do you mean that you'd try to partition tables
(e.g. - if there's a Seq Scan on a table with 8 1GB segments, you
could spawn as many as 8 processes), or that two scans that are then
merge joined means a process for each scan, and a process for the
merge join? Or perhaps both :-). Or perhaps something else entirely ;-).
> To take care of the startup latency problem, we persist sets of
> these backends, called "gangs". They appear, persist for connection
> scope for reuse, then are disbanded.
If only that could happen to more gangs...
--
output = ("cbbrowne" "@" "cbbrowne.com")
http://cbbrowne.com/info/multiplexor.html
"I'm sorry, the teleportation booth you have reached is not in service
at this time. Please hand-reassemble your molecules or call an
operator to help you...."