Re: Improving hvm IO performance by using self IO emulator (YA io-emu?)

[Anthony Liguori <aliguori@us.ibm.com>]

tgingold@free.fr wrote:
>> KVM has a pretty much optimal path from the kernel to userspace.  The
>> overhead of going to userspace is roughly two syscalls (and we've
>> measured this overhead).  Yet it makes almost no difference in IO
>> throughput.
>>     
> The path can be split into 2 parts: from trap to ioemu and from ioemu to
> real hardware (the return is the same).  ioemu to hardware should be roughly
> the same with KVM and Xen.  Is trap to ioemu that different between Xen and
> KVM ?
>   

Yup.  With KVM, there is no scheduler involvement.  qemu does a blocking 
ioctl to the Linux kernel, and the Linux kernel does a vmrun.  Provided 
the time slice hasn't been exhausted, Linux returns directly to qemu 
after a vmexit.

Xen uses event channels which involved domain switches and 
select()'ing.  A lot of the time, the path is pretty optimal.  However, 
quite a bit of the time, you run into worst case scenarios with the 
various schedulers and the latency sky rockets.

> Honestly I don't know.  Does anyone have figures ?
>   

Yeah, it varies a lot on different hardware.  For reference:

if round trip to a null int80 syscall is 150 nsec, a round trip vmexit 
to userspace in KVM may be 2500 nsec.  On bare metal, it may cost 1700 
nsec to do a PIO operation to a IDE port so 2500 really isn't that bad.

Xen is usually around there too but every so often, it spikes to 
something awful (100ks of nsecs) and that skews the average cost.

> It would be interesting to compare disk (or net) performances between:
> * linux
> * dom0
> * driver domain
> * PV-on-HVM drivers
> * ioemu
>
> Does such a comparaison exist ?
>   

Not that I know of.  I've done a lot of benchmarking but not of PV-on-HVM.

Xen can typically get pretty close to native for disk IO.

>> The big problem with disk emulation isn't IO latency, but the fact that
>> the IDE emulation can only have one outstanding request at a time.  The
>> SCSI emulation helps this a lot.
>>     
> IIRC, a real IDE can only have one outstanding request too (this may have
> changed with AHCI).  This is really IIRC :-(
>   

You recall correctly.  IDE can only have one type of outstanding DMA 
request.

> BTW on ia64 there is no REP IN/OUT.  When Windows use IDE in PIO mode (during
> install and crash dump), performances are horrible.  There is a patch which
> adds a special handling for PIO mode and really improve data rate.
>   

Ouch :-(  Fortunately, OS's won't use PIO very often.

>> I don't know what the bottle neck is in network emulation, but I suspect
>> the number of copies we have in the path has a great deal to do with it.
>>     
> This reason seems obvious.
>
>
> [...]
>   
>> There's a lot to like about this sort of approach.  It's not a silver
>> bullet wrt performance but I think the model is elegant in many ways.
>> An interesting place to start would be lapic/pit emulation.  Removing
>> this code from the hypervisor would be pretty useful and there is no
>> need to address PV-on-HVM issues.
>>     
> Indeed this is the simpler code to move.  But why would it be useful ?
>   

Removing code from the hypervisor reduces the TCB so it's a win.  Having 
it in firmware within the HVM domain is even better than having it in 
dom0 too wrt the TCB.

>> Can you provide more details on how the reflecting works?  Have you
>> measured the cost of reflection?  Do you just setup a page table that
>> maps physical memory 1-1 and then reenter the guest?
>>     
> Yes, set disable PG, set up flat mode and reenter the guest.
> Cost not yet measured!
>   

That would be very useful to measure.  My chief concern would be that 
disabling PG would be considerably more costly than entering with paging 
enabled.  That may not be the case on VT today since there is no ASIDs 
so it would be useful to test on SVM too.

>> Does the firmware get loaded as an option ROM or is it a special portion
>> of guest memory that isn't normally reachable?
>>     
> IMHO it should come with hvmload.  No needs to make it unreachable.
>   

It would be nice to get rid of hvmloader in the long term IMHO.  Any 
initialization should be done in the BIOS.

Regards,

Anthony Liguori

> Tristan.
>