[rrd-users] Multiple Updates Per Second

Graham Leggett minfrin at sharp.fm
Tue Oct 2 23:03:23 CEST 2012


On 02 Oct 2012, at 10:32 PM, Wesley Wyche <wesley.wyche at verizon.com> wrote:

> Is there anyone that has had success in dealing with large numbers of
> updates per second, and if so, what solution are you using?

What I have done in the past is to "collapse" updates on the fly to reduce their physical number, and in turn the number of writes that has to be made to RRD files. What I modelled it on was "sort" followed by "uniq -c", which turned this (3 writes):

/foo
/foo
/bar

into this (2 writes):

2: /foo
1: /bar

Or at scale (584951 writes becomes 2 writes):

487264: /foo
97687: /bar

Obviously "sort" doesn't work on streams, so I came up with a tool that did the "uniq -c" part on streamed data. It kept a FIFO of lines, a line that didn't already exist went in at the head. If the line did exist in the FIFO a counter was incremented. Once a timeout for each line had been reached, a line was ejected. This "collapsed" data on the fly. Increasing this timeout reduced the number of lines, and in turn the number of rrd writes. All of this was processed in real time, which meant I could drop the date out of the line format and just count URLs only, over multiple seconds.

The storage was multiple rrd files, one per URL, and the bottleneck was the number of discrete URLs. Ultimately it didn't matter as to the magnitude of the numbers being written, but rather than physical number of URLs involved. If you use something like "cut" you can turn "/foo/subfoo/subsubfoo" into just "/foo" and then "collapse" that, thus reducing the flood of data to a trickle. You lose detail in the data, but this could be the difference between "possible" and "impossible". One of the key things about this is that to get real performance you need to do this in a language like C. As soon as a scripting language got involved performance went down by orders of magnitude and became the bottleneck.

Depending on the kind of data you're trying to record, rrdcached may or may not help. If you have data being saved across tens or hundreds of GBs of rrd files, then rrdcached is unlikely to help as you need a cache hundreds of GB in size. If you have some hot rrd files then rrdcached might work for you, but it's not guaranteed, it all depends on the data you're recording.

I found iowait to be the key bottleneck, which something like an SSD might help with, assuming you don't wear it out in the process.

Regards,
Graham
--

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4365 bytes
Desc: not available
Url : http://lists.oetiker.ch/pipermail/rrd-users/attachments/20121002/a004e912/attachment.bin 


More information about the rrd-users mailing list