Re: Intel IOMMU (and IOMMU for Virtualization) performances

[mark gross <mgross@linux.intel.com>]

On Wed, Jun 04, 2008 at 11:47:01PM +0900, FUJITA Tomonori wrote:
> I resumed the work to make the IOMMU respect drivers' DMA alignment
> (since I got a desktop box having VT-d). In short, some IOMMUs
> allocate memory areas spanning driver's segment boundary limit (DMA
> alignment). It forces drivers to have a workaround to split up scatter
> entries into smaller chunks again. To remove such work around in
> drivers, I modified several IOMMUs, X86_64 (Calgary and Gart), Alpha,
> POWER, PARISC, IA64, SPARC64, and swiotlb.
> 
> Now I try to fix Intel IOMMU code, the free space management
> algorithm.
> 
> The major difference between Intel IOMMU code and the others is Intel
> IOMMU code uses Red Black tree to manage free space while the others
> use bitmap (swiotlb is the only exception).
> 
> The Red Black tree method consumes less memory than the bitmap method,
> but it incurs more overheads (the RB tree method needs to walk through
> the tree, allocates a new item, and insert it every time it maps an
> I/O address). Intel IOMMU (and IOMMUs for virtualization) needs
> multiple IOMMU address spaces. That's why the Red Black tree method is
> chosen, I guess.
> 
> Half a year ago, I tried to convert POWER IOMMU code to use the Red
> Black method and saw performance drop:
> 
> http://linux.derkeiler.com/Mailing-Lists/Kernel/2007-11/msg00650.html
> 
> So I tried to convert Intel IOMMU code to use the bitmap method to see
> how much I can get.
> 
> I didn't see noticable performance differences with 1GbE. So I tried
> the modified driver of a SCSI HBA that just does memory accesses to
> emulate the performances of SSD disk drives, 10GbE, Infiniband, etc.
> 
> I got the following results with one thread issuing 1KB I/Os:
> 
>                     IOPS (I/O per second)
> IOMMU disabled         145253.1 (1.000)
> RB tree (mainline)     118313.0 (0.814)
> Bitmap                 128954.1 (0.887)
>

FWIW: You'll see bigger deltas if you boot with intel_iommu=strict, but
those will be because of waiting on IOMMU hardware to flush caches and
may further hide effects of gong with a bitmap as opposed to a RB tree.
 
> 
> I've attached the patch to modify Intel IOMMU code to use the bitmap
> method but I have no intention of arguing that Intel IOMMU code
> consumes more memory for better performance. :) I want to do more
> performance tests with 10GbE (probably, I have to wait for a server
> box having VT-d, which is not available on the market now).
> 
> As I said, what I want to do now is to make Intel IOMMU code respect
> drivers' DMA alignment. Well, it's easier to do that if Intel IOMMU
> uses the bitmap method since I can simply convert the IOMMU code to
> use lib/iommu-helper but I can modify the RB tree method too.
>

I'm going to be out of contact for a few weeks but this work sounds
interesting.  
 
> I'm just interested in other people's opinions on IOMMU
> implementations, performances, possible future changes for performance
> improvement, etc.
> 
> For further information:
> 
> LSF'08 "Storage Track" summary by Grant Grundler:
> http://iou.parisc-linux.org/lsf2008/SUMMARY-Storage.txt
> 
> My LSF'08 slides:
> http://iou.parisc-linux.org/lsf2008/IO-DMA_Representations-fujita_tomonori.pdf
> 
> 
> Tis patch is against the latst git tree (note that it just converts
> Intel IOMMU code to use the bitmap. It doesn't make it respect
> drivers' DMA alignment yet).
> 

I'll look closely at your patch later.

--mgross