C110 builtin nic slow?
Hi all,
Here are the devices found on my c110 box:
Found devices:
1. U2-IOA BC Runway Port (12) at 0xfff88000 [8], versions 0x580, 0x7, 0xb
2. SkyHawk 100/120 FW-SCSI (4) at 0xf3f8c000 [8/12], versions 0x1f, 0x0, 0x89
3. Raven T' Core BA (11) at 0xffd00000 [8/16], versions 0x32, 0x0, 0x81,
additional addresses: 0xffd0c000 0xffc00000
4. Raven T' Core Centronics (10) at 0xffd02000 [8/16/0], versions 0x32, 0x0, 0x74, additional addresses: 0xffd01000 0xffd03000
5. Raven T' Audio (10) at 0xffd04000 [8/16/1], versions 0x32, 0x0, 0x7b
6. Raven T' Lasi Core RS-232 (10) at 0xffd05000 [8/16/4], versions 0x32, 0x0, 0x8c
7. Raven T' Core SCSI (10) at 0xffd06000 [8/16/5], versions 0x32, 0x0, 0x82
8. Raven T' Core LAN (802.3) (10) at 0xffd07000 [8/16/6], versions 0x32, 0x0, 0x8a
9. Raven T' Core PS/2 Port (10) at 0xffd08000 [8/16/7], versions 0x32, 0x0, 0x84
10. Raven T' Core PS/2 Port (10) at 0xffd08100 [8/16/8], versions 0x32, 0x0, 0x84
11. Raven T' Core PC Floppy (10) at 0xffd0a000 [8/16/10], versions 0x32, 0x0, 0x83
12. Raven T' Wax BA (11) at 0xffe00000 [8/20], versions 0x1e, 0x0, 0x8e,
additional addresses: 0xffe03000 0xffe06000
13. Raven T' Wax HIL (10) at 0xffe01000 [8/20/1], versions 0x1e, 0x0, 0x73
14. Raven T' Wax RS-232 (10) at 0xffe02000 [8/20/2], versions 0x1e, 0x0, 0x8c
15. Raven T' Wax EISA BA (11) at 0xfc000000 [8/20/5], versions 0x1e, 0x0, 0x90, additional addresses: 0xffc88400 0xf4000000
16. U2-IOA BC GSC+ Port (7) at 0xf3fbf000 [8/63], versions 0x501, 0x1, 0xc, additional addresses: 0xf3f80000
17. U2-IOA BC Runway Port (12) at 0xfff8a000 [10], versions 0x580, 0x7, 0xb
18. Raven T' GSC Core Graphics (10) at 0xf4000000 [10/16], versions 0x32, 0x0, 0x85, additional addresses: 0xf0069000
19. U2-IOA BC GSC+ Port (7) at 0xf3fff000 [10/63], versions 0x501, 0x1, 0xc
20. Raven 120 T' (0) at 0xfffa0000 [32], versions 0x58e, 0x0, 0x4
21. Memory (1) at 0xfffb1000 [49], versions 0x49, 0x0, 0x9
CPU(s): 1 x PA7200 (PCX-T') at 120.000000 MHz
And, at the office, I use to obtain around nice 1Mb/s when I do such rsync or ftp between my b180 connected via a hub.
But here at home, I connect the builtin nic of my c110 with a cross-cable to my pc (proxy) and I just obtain 50kb/s (whatever the kernel I boot 2.4 or 2.6) (ie 20 time less then with a another pc connected in place of the c110). Is it possible to improve the performance of this builtin nic?
Thanks in advance,
Joel
Quoted 2 lines · Joel Soete
8. Raven T' Core LAN (802.3) (10) at
0xffd07000[8/16/6], versions0x32,0x0,0x8a
This is a 10Mb link.
Quoted 2 lines
And, at the office, I use to obtain around nice 1Mb/s when I do such rsync or ftp between my b180 connected via a hub.
Which protocol?
Quoted 5 lines
But here at home, I connect the builtin nic of my c110 with a cross-cable to my pc (proxy) and I just obtain 50kb/s (whatever the kernel I boot 2.4 or 2.6) (ie 20 time less then with a another pc connected in place of the c110). Is it possible to improve the performance of this builtin nic?
Earlier this year I exchanged email with someone on linux-ia64 list with a similar sounding problem. They were trying to NFS mount some exported by the ia64-linux server but perf was ~80KB/s vs 50MB/s (1000BT) across the same link to another ia64-linux box. We never found the root cause since netperf demonstrated the UDP throughput was > 50MB/s (expected) in the same config. I could only guess it was something in the NFS stack having to do with 16K pages.
But thinking about this more, I'm wondering if some kernel code is accessing misaligned data someplace in the networking stack. The arguments I've heard is this is expected behavior.
Joel,
1) Can you verify dmesg has no misaligned data access reports?
2) Can you clarify how you are measuring performance?
(ie which protocols and which tools?)
3) Can you setup/run netperf or httperf with the PC to verify whatever protocol you are using basically works?
grant
Hi Grant,
Quoted 7 lines · Grant Grundler
Quoted 1 line · Joel Soete
8. Raven T' Core LAN (802.3) (10) at
0xffd07000[8/16/6], versions0x32,0x0,0x8aThis is a 10Mb link.
Yes ;)
Quoted 6 lines
Quoted 2 lines
And, at the office, I use to obtain around nice 1Mb/s when I do such rsync or ftp between my b180 connected via a hub.
Which protocol?
ftp (ncftp), rsync, scp
Quoted 4 lines
Quoted 2 lines
But here at home, I connect the builtin nic of my c110 with a cross-cable to my pc (proxy) and I just obtain 50kb/s (whatever the kernel I boot 2.4 or 2.6) (ie 20 time less then with a another pc connected in place of the c110). Is it possible to improve the performance of this builtin nic?
Quoted 8 lines
Earlier this year I exchanged email with someone on linux-ia64 list with a similar sounding problem. They were trying to NFS mount some exported by the ia64-linux server but perf was ~80KB/s vs 50MB/s (1000BT) across the same link to another ia64-linux box. We never found the root cause since netperf demonstrated the UDP throughput was > 50MB/s (expected) in the same config. I could only guess it was something in the NFS stack having to do with 16K pages.
Well all kernel tested since now (2.4 or 2.6) were well configured to support NFS (even if exports file is empty). I can try to remove this support?
Quoted 6 lines
But thinking about this more, I'm wondering if some kernel code is accessing misaligned data someplace in the networking stack. The arguments I've heard is this is expected behavior.
Joel,
1) Can you verify dmesg has no misaligned data access reports?
Unfortunately no such a messages :(
Quoted 2 lines
2) Can you clarify how you are measuring performance? (ie which protocols and which tools?)
what ever the protocol I use (ftp, rsync, scp) they report their own stat which are very low and correspond to the results of gkrellm (iptraf seems to lock the trafic).
hm another strange thing: I just compressed the my last kernel 2.4 sources (about 30Mb) and on my pc I get it with scp from the c110: scp and gkrellm showing a rate of about 150Kb/s. Then remove the file on the c110 and on this i grab the same file with scp (also) and this time scp amd gkrellm showing togehter a rate a about 50Kb/s. That is get now put show same results (i mean on pc put to c110: rate about 50Kb/s; on c110 put to pc: rate about 150Kb/s).
That say. I also try to re-do the same test with ncftp which seems to show the same results excepted that 'put' from c110 to pc 'stail' after only one hundred Kb (and I installed the same ftp server on each ie:
ii ftpd 0.17-16 FTP server ii tftpd 0.17-10 Internet trivial file transfer protocol serv).
Quoted 2 lines
3) Can you setup/run netperf or httperf with the PC to verify whatever protocol you are using basically works?
I try to install the non-free (?) dpkg on my pc but:
sid:/home/jso/work # netperf -t UDP_STREAM UDP UNIDIRECTIONAL SEND TEST to localhostudp_send: data send error: Message too long sid:/home/jso/work # netperf -t UDP_STREAM -f m UDP UNIDIRECTIONAL SEND TEST to localhostudp_send: data send error: Message too long sid:/home/jso/work # netperf -t TCP_RR
TCP REQUEST/RESPONSE TEST to localhost
Local /Remote
Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec
16384 87380 1 1 9.99 15980.06
16384 87380
sid:/home/jso/work # netperf -t UDP_RR
UDP REQUEST/RESPONSE TEST to localhost
Local /Remote
Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec
65535 65535 1 1 10.00 27145.26
65535 65535
hmm may be should I try to rebuild the very last src (from hp).
More over the pkg is not available for parisc but I trust it would just be rebuild for this platform?
(i just haven't enough time to do it right now, sorry)
The few I can let iptraf running it didn't show's me any udp trafic :(
Statistics for eth1
Total: 5350 4077724 2411 459794 2939 3617930 IP: 5350 4002824 2411 426040 2939 3576784 TCP: 5350 4002824 2411 426040 2939 3576784 UDP: 0 0 0 0 0 0 ICMP: 0 0 0 0 0 0 Other IP: 0 0 0 0 0 0 Non-IP: 0 0 0 0 0 0
Thanks for help,
Joel
Hi Grant,
Sorry to answer to my mail but
Quoted 4 lines
That say. I also try to re-do the same test with ncftp which seems to show the same results excepted that 'put' from c110 to pc 'stail' after only one hundred Kb (and I installed the same ftp server on each ie:
ii ftpd 0.17-16 FTP server ii tftpd 0.17-10 Internet trivial file transfer protocol serv).
How may I find and reset the speed setup of a nic (I mean which 10 or 100 Mb, half or full duplex)?
In a second though I just remember that the pc boot first (i need minicom console to boot c110) then latter the c110. So it could be that the first one is set in 100 half as default (no connection at boot time) and the second 10 half?
Thanks,
Joel
Quoted 2 lines · Joel Soete
How may I find and reset the speed setup of a nic (I mean which 10 or 100 Mb, half or full duplex)?
ethtool?
Quoted 4 lines
In a second though I just remember that the pc boot first (i need minicom console to boot c110) then latter the c110. So it could be that the first one is set in 100 half as default (no connection at boot time) and the second 10 half?
shouldn't matter unless auto-negotiation is broken for one driver or the other. If they are point-to-point, I expect the negotiation to take place when the link is brought up with the second NIC.
grant
Quoted 5 lines · Grant Grundler
Quoted 2 lines · Joel Soete
How may I find and reset the speed setup of a nic (I mean which 10 or 100 Mb, half or full duplex)?
ethtool?
I know the lasi_82596 driver doesn't support ethtool yet.
Someone who was interested and had time & the right machine could do some serious work on lasi_82596:
- Get rid of "static struct net_device *netdevs[MAX_DRIVERS];"
- Make it work on peecees and 680x0 too and then get rid of the original 82596 driver.
- Add ethtool support.
Signature · 3 lines
"It's not Hollywood. War is real, war is primarily not about defeat or victory, it is about death. I've seen thousands and thousands of dead bodies. Do you think I want to have an academic debate on this subject?" -- Robert Fisk
Grant, Matthew,
Quoted 8 lines
Quoted 5 lines · Joel Soete
How may I find and reset the speed setup of a nic (I mean which 10 or
100 Mb, half or full duplex)?
ethtool?
I will check on my own pc (hp cards but don't remember the model)
hmm I test here with a 3C905B (on an hp vectra) but:
# ethtool eth0 Settings for eth0: No data available
???
(same results on a b180 with eth0: Digital DS21143 Tulip rev 65 [last kernel 2.4.21])
Quoted 9 lines
Quoted 6 lines
In a second though I just remember that the pc boot first (i need minicom console to boot c110) then latter the c110. So it could be that
the first one is set in 100 half as default (no connection at boot time)
and the second 10 half?
shouldn't matter unless auto-negotiation is broken for one driver or the other. If they are point-to-point,
It is a cross-cable
Quoted 3 lines
I expect the negotiation to take place when the link is brought up with
the second NIC.
I am curious to see if ethtool works for my pc's nic to check the behaviour on this side.
Thanks for all info,
Joel
-------------------------------------------------------------------------
Tiscali ADSL: 3 mois GRATUITS! L'Internet rapide, c'est pour tout le monde.
Hi all,
Sorry for this auto-answer but don't see how better follow-up without breacking thread?
Quoted 6 lines · Joel Soete
Grant, Matthew,
I will check on my own pc (hp cards but don't remember the model)
hmm I test here with a 3C905B (on an hp vectra) but:
Quoted 9 lines
# ethtool eth0 Settings for eth0: No data available???
(same results on a b180 with eth0: Digital DS21143 Tulip rev 65 [last kernel 2.4.21])
And unfortunately no better results:
[...]
hp100: eth0: Busmaster mode enabled.
hp100: eth0: HP J2585B at 0xe400, IRQ 10, PCI bus, 32k SRAM (rx/tx 75%).
hp100: eth0: Adapter is attached to 10Mb/s network.
[...]
hmm I don't find the second nic in may dmesg (I just know that is also an hp but a bit older)?
Anyway:
# ethtool eth1 Settings for eth1: No data available # ethtool eth0 Settings for eth0: No data available
Any idea (may be switching the two interfaces: I mean just change interfaces)?
Thanks again for your attention,
Joel
Quoted 10 lines · Joel Soete
Anyway:
# ethtool eth1 Settings for eth1: No data available # ethtool eth0 Settings for eth0: No data availableAny idea (may be switching the two interfaces: I mean just change interfaces)?
As I said, the lasi_82596 driver doesn't support ethtool.
Signature · 3 lines
"It's not Hollywood. War is real, war is primarily not about defeat or victory, it is about death. I've seen thousands and thousands of dead bodies. Do you think I want to have an academic debate on this subject?" -- Robert Fisk
Quoted 1 line
As I said, the
lasi_82596driver doesn't support ethtool.
My bad: confusing lasi and lasi_82596; sorry for my misunderstanding and thanks for clarification ;)
Joel
-------------------------------------------------------------------------
Tiscali ADSL: 3 mois GRATUITS! L'Internet rapide, c'est pour tout le monde.
Quoted 1 line · Joel Soete
hmm may be should I try to rebuild the very last src (from hp).
netperf home is netperf.org.
Quoted 2 lines
More over the pkg is not available for parisc but I trust it would just be rebuild for this platform?
It does run/build on parisc. I've used it before.
Note that with B180, I can only get about 1MB/s throughput max. With C3000 I get about 5 or 6 MB/s 100BT link.
Your network performance numbers are pretty far off from that but there might be other factors involved.
I expect mem copy routines are saturating the CPU and thrashing the data cache. The tulip driver copies the entire buffer from where the card DMA'd to the skbuff in order to "bias" the buffer by 2 bytes. This avoids un-aligned data access in the "common" TCP case. Adding "unaligned access" macros to the "common" networking code path has been discussed and rejected (again) in the past month or so. Linus himself has rejected such proposals in the past too.
If you twiddle the tulip driver to NOT copy all the data for packets > 512 bytes, the kernel will tell you when/where the network stack makes unaligned accesses. You can fix those and keep that as a local patch. I'd be interested in hearing the netperf (or httperf) numbers with such a patch when compared to the current behavior.
thanks,
grant