blynch at sharpcast.com
Thu Apr 10 22:37:39 UTC 2008
I made the change to using lighttpd for reads across two of the storage
nodes and already see a reduction in the number of timeout errors.
There appears to be an unconfirmed improvement in the speed of our
application as well. I'm letting it burn in a bit before rolling to the
remaining nodes, but things look good so far. Thanks again for the
rapid response and great suggestions!
From: dormando [mailto:dormando at rydia.net]
Sent: Thursday, April 10, 2008 2:42 PM
To: Mark Smith
Cc: Brian Lynch; mogilefs at lists.danga.com
Subject: Re: Mogstored Tuning
You can get a little bit of an idea because a mogstored is essentially a
perlbal plugin. So the management interface works, and you can write a
little extra code to track stats if you wanted.
Otherwise a general idea should be fine, or just purely a reqs/sec with
an understanding of how many writes you're probably doing.
Mark Smith wrote:
>> We are using mogstored for both reads and writes. I wasn't aware
>> the ability to split out reads, but I found the command in mogadm.
>> give that a spin. Thanks for the suggestion!
> Definitely definitely do that, that's one of the first things we did
> when we actually started using MogileFS in a serious way on LJ.
> Apache2 works, lighttpd works, whatever can do GETs in a quick way
> will work!
>> Is there a way to gather statistics on the number of reads/writes
>> passing through mogstored?
> Not built in, but you should be able to determine this by just
> understanding your traffic. Unless you are overwriting files
> constantly, or serving as a backup service where people don't get
> their data often, there's only a few writes per file. One to insert,
> N to replicate up to mindevcount, and then many reads for serving.
> The files aren't touched again (unless you do something with fsck or
> rebalancing, but even then you're still heavily on the side of reads
> for most usage).
More information about the mogilefs