hast vs ggate+gmirror sychrnoisation speed

[Mikolaj Golub]

On Thu, 28 Oct 2010 18:30:36 +0200 Pawel Jakub Dawidek wrote:

 PJD> On Wed, Oct 27, 2010 at 10:05:20PM +0300, Mikolaj Golub wrote:
 >> In hast_proto_send() we send header and then data. Couldn't it be that
 >> remote_send and sync threads interfere and their packets are mixed? May be some
 >> synchronization is needed here?
 >> 
 >> I set sleep(1) in hast_proto_send() between proto_send(header) and
 >> proto_send(data). The error started to occur frequently.

 PJD> Synchronization requests are sent through the remote thread just like
 PJD> regular I/O requests, exactly because of races that can occur.

 PJD> I looked at the code and the keepalive packets arbe sent from another
 PJD> thread. Could you try turning them off in primary.c and see if that
 PJD> helps?

At first I set RETRY_SLEEP to 1 sec to have more keepalive packets. The errors
started to observe frequently:

Oct 28 21:35:53 bolek hastd[1709]: [storage] (secondary) Unable to receive request header: RPC version wrong.
Oct 28 21:35:54 bolek hastd[1632]: [storage] (secondary) Worker process exited ungracefully (pid=1709, exitcode=75).
Oct 28 21:36:12 bolek hastd[1722]: [storage] (secondary) Unable to receive request header: RPC version wrong.
Oct 28 21:36:12 bolek hastd[1632]: [storage] (secondary) Worker process exited ungracefully (pid=1722, exitcode=75).
...

Now I have been running synchronization for more then a half an hour with
keepalive_send disabled and have not seen any error.

-- 
Mikolaj Golub