myrinet 2000 athlon transfer rate
Mark Wiederspahn <[email protected]>
| Newsgroups | gmane.network.myrinet.general |
|---|---|
| Message-ID | <[email protected]> |
Dear myrinet list,
I'm testing gm-1.5 on RedHat 7.1 AMD dual athlon tyan motherboard with
M3M-PCI64B-2 boards and a 16 port switch against our existing
alpha 164lx cluster.
gm_debug -L says in part:
DMA rate for 4096 Byte transfers (64bit / 33MHz bus)
bus_read (send) = 240 MBytes/s
bus_write (recv) = 204 MBytes/s
vs. our old alpha lx164 systems:
DMA rate for 8192 Byte transfers (64bit / 33MHz bus)
bus_read (send) = 109 MBytes/s
bus_write (recv) = 240 MBytes/s
HOWEVER, the actual limiting bandwidth of alpha-alpha when
moving data seems to be faster, at 100 Mbytes/sec vs.
80 Mbytes/sec going athlon to alpha, using logp_test.
I don't understand this result. We had measured some cheap 32bit
pci athlons, which were disappointing at logp_test = 40ish, but
had assumed the problem was just pci dma, not other more
complicated issues.
DMA rate for 4096 Byte transfers (32bit / 33MHz bus)
1st: 8 pages from bogus sdma pg, 8 to bogus rdma
bus_read (send) = 63 MBytes/s
bus_write (recv) = 97 MBytes/s
(We yet don't have two dual-cpu athlons in house to try the
obvious simple test of dual athlon-dual athlon.)
thanks for any light,
mw
-- mailto:[email protected]
Mark Wiederspahn Senior System Analyst
Inst. for Geophysics University of Texas at Austin
4412 Spicewood Springs Rd. #600
Austin TX 78759-8500 phone: (512)471-0406
fax: (512)471-8844
http://www.ig.utexas.edu