[dm-devel] dd slow while reading on multipathing, writing is fast

[Bart Coninckx]

On 02/09/11 12:22, Bryn M. Reeves wrote:
> On 02/08/2011 02:44 PM, bart.coninckx at telenet.be wrote:
>> Hi all,
>>
>> something peculiar I cannot get my head round. I have a setup where a backup
>> server connects to iSCSI LUNs over multipathing. While dd-ing to it, I get a
>> spiffing 200 MB/sec, while reading it drops to 70 MB/sec. Local storage where
>> I write the image onto is sufficiently fast.
> 
> As Christoph mentioned the versions of the kernel and tools you are using and
> the configuration of the multipath device would be useful but there are a couple
> of things that spring to mind; if you are seeing good performance on write but
> poor on read then it's possible that your testing is not really measuring the
> device performance.
> 
> Writes are buffered in memory (possibly both on the iSCSI initiator and target,
> depending on configuration) whereas a read of an un-cached region of the device
> will need to perform actual I/O before it can return. Although dd is a useful
> tool for quick tests it's not a very rigorous way to benchmark performance
> unless you take special steps to mitigate things like caching effects.
> 
> Another point is that you don't mention if you have tested the underlying iSCSI
> device in isolation? I.e. remove multipath from the picture and verify that you
> get the kind of performance you expect from a single path.
> 
> Regards,
> Bryn.

All,


This is on Opensuse 11.3 with kernel-desktop-2.6.34.7-0.7.1.x86_64
(probably should change that to a default kernel). I upgraded the
multipahting-tools to the latest version because they are badly broken
on the 64-bit version of this distro.

I used bs and count combinations for the dd command that result into a
dump that surpasses available RAM, so there should be no caching happening.
I later on did indeed try on iscsi without multipathing on a bonded
interface and got about the same results, but I have to admit that I
forgot to change tc_reordering, which is necessary on Suse. probably
better do that again and report back. I did find however that if the
underlying iscsi device is an LVM snaphost, read speed goes down when
LVM snapshot size goed up (or so it seems).

b.