RE: [Xen-users] GPL PV drivers for Windows 0.9.11-pre12

["James Harper" <james.harper@xxxxxxxxxxxxxxxx>]

> 
> I've installed GPL PV drivers in XP64 using 2k3x64 check of 0.9.11-pre13
> and everything went fine (functionally)...

Thanks for the report.

> Appart from that, the
> performance was actually worst than using the default qemu devices...
> 
> I've already sent a message with the config i used but, briefly: dom 0
> 3x500 - raid5 - lvm partitions for VM disks. 40Mb/s in dom0, 15Mb/s in
> xp64 domU with qemu and 4Mb/s enabling /gplpv...
> 
> Any clue?
> 

No, but there are enough of these reports around that I'm making it a priority 
to look at next.

Can you tell me what tools you are using for testing?

I know that there is a performance related problem with buffer alignment, 
although I'm not sure how it relates to real world performance, or if it 
relates at all to the poor performance that people have been seeing while 
benchmarking.

Basically, Linux always aligns block device request buffers to a 512 byte 
(sector) boundary in memory, eg the address of the buffer will always be 
divisible by 512. The Dom0 xen block device driver backend enforces this too, 
so I have to comply.

Windows, on the other hand, doesn't care, and only mostly aligns request 
buffers to a 512 byte boundary, so I have to double buffer the requests (eg 
copy to (for a write) or from (for a read) a secondary buffer that is aligned 
to a 512 byte boundary. There are small overheads with this but the main 
problem is that the secondary buffer is only small and so a large read or write 
request gets broken down into several requests that must be executed serially, 
eg part 2 of the request can only be submitted once part 1 has been completed. 
Fortunately I have only seen Windows issue large numbers of non-aligned 
requests during format, chkdsk (online and boot) and certain benchmarking tools.

A disk driver under Windows using the scsiport framework (eg xenvbd) executes 
at a very high IRQ Level, and it is impossible to allocate memory at that IRQL 
so I'm limited as to what buffers I can make use of.

Sorry you asked? :)

I really would like to know of a simple, easy, and fast tool (eg command line 
based) to do testing under a Linux DomU and a Windows DomU. It doesn't have to 
be incredibly accurate, as long as I can make comparisons between the two 
environments... I find iometer painful :(

Thanks

James


_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users