[kernel-xen] kernel-xen-3.11.0-1 in testing repo.

[Glenn Enright]

On 09/17/13 13:56, Steven Haigh wrote:
> On 17/09/13 11:30, Glenn Enright wrote:
>> Preliminary (very) testing seems to indicate a good boost (using 3.11.1)
>>
>> Seeing about a 20% increase on buffered writes, and a small but marked
>> increase on syncronised writes. Which is nice. Keep in mind this is on
>> an unloaded host with only a couple of VMs.
>>
>> In detail ...
>>
>> I'm working on a machine with VMs backed by lvm disk over raid1 (using
>> budget sata drives), with the following snippet, with and without the
>> fsync option. The test domU uses the noop scheduler.
>>
>> for i in 1 2 3 4 5
>> do
>>    dd if=/dev/zero of=speedtest bs=1M count=1024 conv=fsync
>> done
>>
>> 3.10 domO, default behavour, custom domu kernel
>> using fsync average 58 MB/s, without fsync average 59 MB/s,
>>
>> 3.11 domO, default behaviour, 3.11 kernel on domu
>> using fsync average 58 MB/s (ie the same)
>>
>> 3.11 domO, using xen-blkfront.max=128, 3.11 kernel on domU
>> using fsync average 62 MB/s, without fsync average 68 MB/s
>
> This is good to see. I'd actually expect a bit more than that - but
> either way, its good to see an increase in throughput. What actual
> drives are these? I use Seagate 1Tb 7200rpm SATA drives in my home test
> setup - but I haven't done a throughput test on the RAID1 as yet - as it
> hasn't really been problematic in the past.
>
> There is talk around this having a way to automatically do a test and
> set xen-blkfront.max to an optimal value on initialisation of the DomU
> blkfront module. As yet, this is only talk - however it would probably
> be a good step forwards.
>
> It may even get to the point where it could be backported into distro
> kernels - and therefore not require the upgrade to 3.11 on the DomU to
> take advantage of this - however that would be a job for various distros.
>
>
>

The drives are pair of WD RE3 250G, nothing special there.

I managed to get the write result up to about 72MB/s average by tweaking 
the read ahead buffer for the raid1. Thats probably about as good as it 
gets given the host OS gets no more than 78Mb/s average on the same pair.

As time permits I'll try with some different numbers and test raid6 and 
as well.