On Wed, Apr 22, 2015, at 02:27 PM, Ciro Iriarte wrote:
Interesting, is the use of several instances needed because cyrus cannot scale with threads in a single instance scenario?
There are two interesting reasons:
1) global locks. There are some - mailboxes.db for example. If you have multiple instances on a single machine, then a lock never blocks up the entire machine.
2) replication and load spreading - right now there's no support for "partial replica" - a Cyrus instance replicates every mailbox to its replica.
The second one is the kicker.
If we replicated everything from one machine to another machine, then we'd have 100% user load on one machine and nothing on the other - not efficient use of resources, because the second one needs to have the capacity to run at 100% in a failover situation too.
Our first thought was to run two instances per machine and pair them - so there was a master on one and a replica on the other. At least then we're running equally in the general situation, and only in a failover situation are we loaded 100%. But it's still nasty - you go from 50% load to 100% load.
So we have about 10 different replicas for each machine, and every machine is running at 50% capacity. If we need to take one machine down, then 10 other machines run at 55% capacity instead for that time. The load change is much less.
(as of about a year ago, we're fully paired odd-host-number to even-host-number, and odd and even are in different cabinets, so we can shut down an entire cabinet by raising the load on its replicas)
Bron.
--
Bron Gondwana
brong@xxxxxxxxxxx
---- Cyrus Home Page: http://www.cyrusimap.org/ List Archives/Info: http://lists.andrew.cmu.edu/pipermail/info-cyrus/ To Unsubscribe: https://lists.andrew.cmu.edu/mailman/listinfo/info-cyrus
