EtherExpressPro & linux-2.2.14

[Steffen Persvold]

The only thing that isn't too good with the Intel driver from their
website is that it's not good at channel bonding (which might interrest
you beowulfers) !

I've performed benchmarks with Netpipe and it shows that the e100 driver
may perform as well as the eepro100 driver in single mode, but if I
'bond' two of them together the eepro100 is by far better. The results I
get a peak performance of only 55 Mbit/sec (!?!) with the e100 driver,
while the eepro100 reach 179 Mbit/sec. The tests were run with two
identical Serverworks 370DER motherboards using the two onboard network
controllers. The nodes was connected to a Intel 460T switch. The kernel
I was using was a patched up 2.2.16 kernel (the bonding driver from
2.2.18)

Some of the output from NPtcp:
eepro100:
   # ./NPtcp -t -h 192.168.1.4 -l 131072 -u 4194307 -P
   Latency: 0.000068
   Now starting main loop
     0:    131069 bytes    7 times -->  172.14 Mbps in 0.005809 sec
     1:    131072 bytes   21 times -->  172.20 Mbps in 0.005807 sec
     2:    131075 bytes   21 times -->  172.18 Mbps in 0.005808 sec
     3:    196605 bytes   21 times -->  174.33 Mbps in 0.008604 sec
     4:    196608 bytes   19 times -->  174.32 Mbps in 0.008605 sec
     5:    196611 bytes   19 times -->  174.31 Mbps in 0.008606 sec
     6:    262141 bytes   19 times -->  175.33 Mbps in 0.011407 sec
     7:    262144 bytes   16 times -->  175.36 Mbps in 0.011405 sec
     8:    262147 bytes   16 times -->  175.36 Mbps in 0.011406 sec
     9:    327677 bytes   10 times -->  176.42 Mbps in 0.014171 sec
    10:    327680 bytes   10 times -->  176.43 Mbps in 0.014170 sec
    11:    327683 bytes   10 times -->  176.42 Mbps in 0.014171 sec
    12:    458749 bytes   10 times -->  177.17 Mbps in 0.019755 sec
    13:    458752 bytes    9 times -->  177.18 Mbps in 0.019754 sec
    14:    458755 bytes    9 times -->  177.18 Mbps in 0.019755 sec
    15:    589821 bytes    7 times -->  177.74 Mbps in 0.025318 sec
    16:    589824 bytes    7 times -->  177.74 Mbps in 0.025318 sec
    17:    589827 bytes    7 times -->  177.70 Mbps in 0.025323 sec
.............

e100:
   # ./NPtcp -t -h 192.168.1.4 -l 131072 -u 4194307 -P -o e100_bond.txt
   Latency: 0.000072
   Now starting main loop
     0:    131069 bytes    7 times -->   54.29 Mbps in 0.018421 sec
     1:    131072 bytes    7 times -->   52.13 Mbps in 0.019184 sec
     2:    131075 bytes    7 times -->   52.51 Mbps in 0.019046 sec
     3:    196605 bytes    7 times -->   52.31 Mbps in 0.028677 sec
     4:    196608 bytes    7 times -->   51.44 Mbps in 0.029160 sec
     5:    196611 bytes    7 times -->   51.97 Mbps in 0.028864 sec
     6:    262141 bytes    7 times -->   51.48 Mbps in 0.038849 sec
     7:    262144 bytes    7 times -->   52.08 Mbps in 0.038402 sec
     8:    262147 bytes    7 times -->   52.21 Mbps in 0.038309 sec
     9:    327677 bytes    7 times -->   51.01 Mbps in 0.049014 sec
    10:    327680 bytes    7 times -->   51.62 Mbps in 0.048434 sec
    11:    327683 bytes    7 times -->   51.19 Mbps in 0.048837 sec
    12:    458749 bytes    7 times -->   51.36 Mbps in 0.068148 sec
    13:    458752 bytes    7 times -->   51.26 Mbps in 0.068283 sec
    14:    458755 bytes    7 times -->   51.21 Mbps in 0.068343 sec
    15:    589821 bytes    7 times -->   51.18 Mbps in 0.087923 sec
    16:    589824 bytes    7 times -->   50.52 Mbps in 0.089071 sec
    17:    589827 bytes    7 times -->   50.78 Mbps in 0.088620 sec


The performance problems with the e100 driver seems to be related to the
receiver; if I use NetPipe in streaming mode (not ping-pong) and use the
e100 driver on the sender and the eepro100 on the receiver performance
is almost back to what it should be.

Conclusion : I think the more permanent solution for everyone using
these adapters must be to see if we can by any chance improve the
stability of the eepro100.

Best regards,
-- 
 Steffen Persvold                        Systems Engineer
 Email  : mailto:sp at scali.com            Scali AS (http://www.scali.com)
 Norway : Tlf  : (+47) 2262 8950         Olaf Helsets vei 6
          Fax  : (+47) 2262 8951         N-0621 Oslo, Norway

 USA    : Tlf  : (+1) 713 706 0544 (43)  Scali
                                         10500 Richmond Avenue, Suite
190
                                         Houston, Texas 77042, USA

DuWayne wrote:
> 
> We had the same problems with the eepro100 diver. We ended up switching to
> Intels driver. Its available from their web site. Cured all the issues we were
> having
> with the eepro100 driver.
> 
> Rob Nelson wrote:
> 
> > >2.2.15 had a much-revised eepro100 driver.
> > >
> > >I am hopeful that using, say, 2.2.18 will cure your problem.
> >
> > FWIW, and I don't really have anything but anecdotal evidence to support
> > this, as the web page at sco.com seems to have disappeared, but replacing
> > eepro100's seems to be the only fix. Take a look at the 2.2.x changelogs.
> > eepro100 drivers have had fatal bugs found in the chipsets multiple times,
> > just in 2.2.x. Our NT, SCO, linux, and even Novell servers almost always
> > have a new eepro driver in whatever upgrade they get. We've finally switched
> > to 3com's in servers due to frustration with Intel at building a quality
> > chipset.
> >
> > Sco used to have a good page detailing the problem. Basically, something on
> > the PCI card that controlled input and output overflowed its onboard 8k ram
> > chip or whatever sort of memory it had, which then killed its internal
> > stack. Power-cycling was the only way to flush it. The bug has been squashed
> > multiple times, only to rear its head in some new form later. If you keep
> > going with eepro's, good luck. I heartily recommend 3c905-b's or -c's.
> >
> > Rob Nelson
> > ronelson at vt.edu
> >
> > _______________________________________________
> > Beowulf mailing list
> > Beowulf at beowulf.org
> > http://www.beowulf.org/mailman/listinfo/beowulf
> 
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf