[OSM-dev] Tuning render_list: beware of hyperthreading and RAID cards latency...

[Christian Quest]

I'm currently doing some rendering benchmark to tune a new tile server
we're about to put online.

A strange thing I've found last night: render_list with 16 thread takes
more time than with 8 .

The server has 2 quad-core Xeon (X5570), this means 8 physical cores. It
looks like hyperthreading adds additionnal context switches or something
similar.
With "-z 9 -Z 9 -n 16 -l 32" it takes 3016s, and with "-z 9 -Z 9 -n 8 -l
12" 2647s, 10% less.

I'll disable HT on my next run in the server bios and see if it is
consistent.


Another thing I found... do not put your SSD behind a RAID card.
The RAID card (Dell Perc 6i) add too much latency.
On the same render_list, the iowait threads are increasing a lot and iops
decreases a lot.

Here is a munin graph clearly showing it: http://cl.ly/image/3V2A1J261o30

Before midnight, the SSD (Samsung 840 EVO 1TB) is on a PCIe/SATA adapter
(Apricorn Velocity Solo x1). After 1:00 it moved to the builtin SAS/SATA
enclosures driven by the RAID card.

The server is a Dell R710 with 48GB RAM running ubuntu server 14.04.
I'm using mapnik 2.3.x with async postgis queries enabled.

-- 
Christian Quest - OpenStreetMap France
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/dev/attachments/20140717/c9da6fb7/attachment.html>