Re: [Postgres-xc-general] pgbench suggestion

[Michael P. <mic...@gm...>]

On Fri, Nov 16, 2012 at 2:12 PM, Tatsuo Ishii <is...@po...> wrote:

> >> >> We have up to 8 servers, where each server has data
> >> >> node/coorrdinator/GTM proxy. We have GTM on a seprate server. Then we
> >> >> execute multiple pgbench -k -S against the servers(each pgbench
> >> >> connect to different server).
> >> >>
> >> > OK, this should not happen...
> >> > How many max connections are you using per Coordinator
> >>
> >> 100.
> >>
> >> > and how many
> >> > Coordinators do you have?
> >>
> >> Up to 4.
> >> The number of coordinators are same as data nodes. So when we test 4
> >> data nodes(suppose they are d1 to d4), we are doing something like
> >> this:
> >>
> >> on host c1: pgbench -c 100 -S -h d1
> >> on host c2: pgbench -c 100 -S -h d2
> >> on host c3: pgbench -c 100 -S -h d3
> >> on host c4: pgbench -c 100 -S -h d4
> >>
> > That looks OK. I was wondering if the maximum active transactions
> possible
> > on GTM (1024) was reached, leading to your error.
> > The warning "Cannot connect to GTM..." "Xid is invalid." only happens
> when
> > a node tries to get a connection from GTM but it cannot, making me think
> > that GTM is not online or you have some firewall/network problems between
> > the node and GTM.
>
> The error seems not depending on number of node. We did the similar
> way as 4 nodes above against 8 nodes configuration. Sometimes we saw
> the error, sometimes not. Also we saw the error when the number of
> nodes less than 4, for example 2 or 1.
>
I believe here in a configuration mistake. I work here with clusters up to
10 servers and do not see that if connection to GTM is made correctly even
with a huge load.
Please double-check gtm_host and gtm_port in each node's postgresql.conf,
then check what is the port open on GTM.
-- 
Michael Paquier
http://michael.otacoo.com