|
From: | Ed W |
Subject: | Re: [Gluster-devel] Can I bring a development idea to Dev's attention? |
Date: | Sun, 26 Sep 2010 01:35:21 +0100 |
User-agent: | Mozilla/5.0 (Macintosh; U; Intel Mac OS X 10.6; en-US; rv:1.9.2.9) Gecko/20100915 Lightning/1.0b2 Thunderbird/3.1.4 |
I would specifically like to optimise the case that you have a flat namespace on the server (master/master filesharing), but you optimise the applications in such a way that the applications running on each brick (NUFA) only touch a subset of all files (in general). eg a mailserver with a flat filesystem, but users are proxied so that they generally touch only a specific server, or a webserver with a flat namespace where a proxy points specific domains to be served by specific servers?
In this case I would like to see a specific brick realise that it's predominantly the reader/write for a subset of all files and optimise it's access at the expense of other bricks which need to access the same files (ie I don't just want to turn up the writeback cache, I want cache coherency across the entire cluster). I would accept that random read/writes to random bricks would be slower, in return for the optimisation that reads/writes would be faster *if* the clients optimise themselves to *prefer* to touch specific bricks (ie NUFA). Such an optimisation should not be set in stone of course, if the activity on a subdirectory generally seems to move across to another brick then that brick should eventually optimise it's read/write performance (at the expense that another brick's access now becomes slower to that same subset of files.)
Anyone care to quote on this? Seems like it's a popular performance issue on the mailing list and with some optimisation later it also seems like the basis for cross datacenter replication?
Thanks Ed W
[Prev in Thread] | Current Thread | [Next in Thread] |