GM DMA performance measurement changes

I notice that somewhere btw 1.4pre22 and pre33 that the code which
calculates the performance of send and receive DMAs in mcp/gmcp.c has
changed.  We were seeing:

    DMA rate for 16384 Byte transfers (64bit / 66MHz bus)
    1st: 16 pages to bogus sdma pg, 16 from bogus rdma
	    bus_read  (send) = 202 MBytes/s
	    bus_write (recv) = 120 MBytes/s

Now the values are:

    DMA rate for 16384 Byte transfers (64bit / 66MHz bus)
    1st: 8 pages from bogus sdma pg, 8 to bogus rdma
	    bus_read  (send) = 195 MBytes/s
	    bus_write (recv) = 95 MBytes/s

Bandwidth numbers for either uni- or bi-directional tests as reported
by gm_allsize went down only a teensy amount, maybe 1%, in spite of the
3% and 20% numbers above.  Latency got better, though, if it's related.
    http://www.osc.edu/~pw/bw-22-33.gif
    http://www.osc.edu/~pw/lat-22-33.gif

Did something major change?  Should I worry?

The reason I noticed this is that in trying to pick which hardware to
buy, I was hoping to use a DMA rate test with myrinet cards to test the
PCI bus.  Can you recommend a way to probe just the bus bandwidth in an
accurate way?  The _old_ myrinet software had some sort of l2e utility
I seem to recall...

		-- Pete