Hello all,
I'm trying to overcome a performance problem we're having with our video
capture card drivers on the G5.
The nature of our problem is that we're seeing drastic differences in
performance between G4 and G5 systems, and not the differences we'd expect.
In a G4 system, when we read from or write to a wired, contiguous, buffer
allocated by a IOBufferMemoryDescriptor call, we see roughly the advertised
memory performance of the system. However on the G5, we are incurring huge
overhead, resulting in significantly slower performance where we would have
expected about a 3X increase.
We have tried literally dozens of things to get around this problem. In our
testing, we have discovered that in User-space, we can allocate two buffers
and copy between them at the expected G5 performance levels. However, if
one of the buffers is our kernel-allocated buffer, it appears we are taking
a huge hit. We've set the buffer up as "user/kernel shared" we've tried
setting the directions to match the direction of the copy. We've even tried
setting the direction as in/out. As an experiment, we've allocated two
buffers in the kernel, and done the copy at interrupt level. In every case,
the performance is an order of magnitude less than the same code on the G4.
Some details:
We allocate wired buffers in the kext, of around 800KBytes (The size of a
standard-definition video frame) each. We allocate them as page-aligned and
contiguous so that we can DMA to/from them from our hardware card without
setting up scatter-gather. This results in almost 200 vm pages per buffer.
We have a QuickTime component which fills these buffers from a given QT
buffer.
At this point, we are assuming that we are incurring a hit brought on by the
DART reloading its address tables, or perhaps issues with VM on the G5. The
same code on G4 machines seems fine.
We have recently come across an Apple soft-VDIG example which includes a
kext, and it looks like our implementation is exactly as in the example, and
as you might expect, the example code performs just as poorly on the G5.
- Mike Stroven
