[DRBD-user] Write Performance: sequencial I/O slower than random I/O

Lars Ellenberg lars.ellenberg at linbit.com
Wed Dec 10 22:37:08 CET 2008

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


On Tue, Dec 09, 2008 at 01:44:37PM +0100, Robert wrote:
> Hello list,
> 
> I'm using DRBD 8.2.7rc2 and doing some benchmarks. The standby machine 
> uses a ramdisk as backend storage to simulate the impact of the network 
> only.
> 
> I also set up different iometer bechmark patterns. For two of them I can 
> not explain the results.
> 
> Pattern1: 100% Write, Blocksize 8k, 100% Random
> Pattern2: 100% Write, Blocksize 8k, 100% Sequencial
> 
> The network between the sites is 100Mbit with a simulated latency of 
> 1500 us +- 500 us.
> 
> DRBD (besides std. config) is:
>   protocol c;
>   al-extents       1021;
>   max-buffers      4096;
>   unplug-watermark 131072;
>   sndbuf-size        0;
> 
> With the same drbd setup I get :
> 
> Pattern1: 685 IO's /Second
> Pattern2: 116 IO's /Second
> 
> I can not explain why sequencial writes are slower than random writes 
> with the same DRBD config.
> 
> I also tried with different unplug-watermark, no effect.
> 
> Any hint ?

watch the network traffic with tcpdump, 
maybe there is some "interessting" timing behaviour there,
congestion control getting in the way or something like that.

play with the max-epoch-size (tune it all the way from the minimum to
the maximum, and plot the benchmark results)
may or may not have any effect.

maybe explicitly set a sndbuf-size.

see if the behaviour is reproduceable for real backing store.

(may be a cpu cache effect, even!)

-- 
: Lars Ellenberg                           http://www.linbit.com :
: DRBD/HA support and consulting             sales at linbit.com :
: LINBIT Information Technologies GmbH      Tel +43-1-8178292-0  :
: Vivenotgasse 48, A-1120 Vienna/Europe     Fax +43-1-8178292-82 :
__
please use the "List-Reply" function of your email client.



More information about the drbd-user mailing list