Loosing packets with Dlink DFE-580TX and SMC 9462TX
Hi,
I'm seeing severe packet-loss with two different types of network cards:
D-link DFE-580TX (DL10050B aka OEM:ed Sundance ST201 chip) and SMC 9462TX (NS83820 chip)
The machine is an Athlon XP1800+ with VIA KT266 chipset and what I'm doing is to send several bi-directional streams of small UDP packets through the machine. Each stream consists of 50 UDP packets per second in each direction, with 200 bytes payload.
Using either Intel 21143 based cards (tulip driver) or Intel 82559 (eepro100 driver) I can send 100 such streams through the machine with no packet loss (the load generators starts walking on their knees a bit over 100). This equals 10000 packets/second or about 19 Mbit/s.
However using either the Dlink cards or the SMC cards I start getting severe packet loss at between 50 and 70 streams. After fixing the sundance driver so it doesn't turn off interrupts for 3.2ms whenever the boguscnt counter in the interrupt handler happens to reach zero I can't find anything particularly bad neither in the driver nor in the ST201 chip spec, but packets keep getting dropped. (The 83820-driver I haven't looked at at all)
Before digging further into this I'd like to know if anyone have had similiar problems, or are using any of these cards/drivers without problems, or have an idea if it is the chips or the drivers which are broken? Or can just tell me whether these cards are exceptionally bad or if eepro100/tulip are exceptionally good?
I migh also take the opportunity to ask if anyone can recommend another 4-port (or more) 100Mbit ethernet card than the DFE-580TX?
//Marcus
Signature · 3 lines
---------------------------------------+-------------------------- Marcus Sundberg <[email protected]> | Firewalls with SIP & NAT Firewall Developer, Ingate Systems AB | http://www.ingate.com/
What version of ns83820.c are you using? Version 0.17 of ns83820.c made significant improvements under load. Other possibilities include cabling problems (watch the kernel logs for changes in link state). Try to find out where the packets are getting dropped by looking through /proc/net/snmp and other statistics counters in the kernel.
-ben
Quoted 44 lines · Marcus Sundberg
Hi,
I'm seeing severe packet-loss with two different types of network cards:
D-link DFE-580TX (DL10050B aka OEM:ed Sundance ST201 chip) and SMC 9462TX (NS83820 chip)The machine is an Athlon XP1800+ with VIA KT266 chipset and what I'm doing is to send several bi-directional streams of small UDP packets through the machine. Each stream consists of 50 UDP packets per second in each direction, with 200 bytes payload.
Using either Intel 21143 based cards (tulip driver) or Intel 82559 (eepro100 driver) I can send 100 such streams through the machine with no packet loss (the load generators starts walking on their knees a bit over 100). This equals 10000 packets/second or about 19 Mbit/s.
However using either the Dlink cards or the SMC cards I start getting severe packet loss at between 50 and 70 streams. After fixing the sundance driver so it doesn't turn off interrupts for 3.2ms whenever the boguscnt counter in the interrupt handler happens to reach zero I can't find anything particularly bad neither in the driver nor in the ST201 chip spec, but packets keep getting dropped. (The 83820-driver I haven't looked at at all)
Before digging further into this I'd like to know if anyone have had similiar problems, or are using any of these cards/drivers without problems, or have an idea if it is the chips or the drivers which are broken? Or can just tell me whether these cards are exceptionally bad or if eepro100/tulip are exceptionally good?
I migh also take the opportunity to ask if anyone can recommend another 4-port (or more) 100Mbit ethernet card than the DFE-580TX?
//Marcus
Signature · 7 lines
---------------------------------------+-------------------------- Marcus Sundberg <[email protected]> | Firewalls with SIP & NAT Firewall Developer, Ingate Systems AB | http://www.ingate.com/ - To unsubscribe from this list: send the line "unsubscribe linux-net" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Signature · 1 line
"You will be reincarnated as a toad; and you will be much happier."
Quoted 5 lines · Benjamin LaHaise
What version of
ns83820.care you using? Version 0.17 ofns83820.cmade significant improvements under load. Other possibilities include cabling problems (watch the kernel logs for changes in link state). Try to find out where the packets are getting dropped by looking through/proc/net/snmpand other statistics counters in the kernel.
0.17, but some more testing showed that the ns83820 actually works just fine during this test when using just crossover cables and running at gigabit speed. The original testing was done using 100Mbit hubs, so my guess would be that the 83820 chips (and/or driver) doesn't handle collisions too well (which I don't have a problem with, as afaik GE is always switched).
However the DFE-580TX problems remain regardless of using a hubbed or switched network.
(As booth eepro100 and tulip-based cards works fine with the hubs I'm quite certain there's nothing wrong with them.)
//Marcud
Signature · 3 lines
---------------------------------------+-------------------------- Marcus Sundberg <[email protected]> | Firewalls with SIP & NAT Firewall Developer, Ingate Systems AB | http://www.ingate.com/
There are at least some gigabit ethernet hubs on the market. How badly does it handle collisions?
Nick
Quoted 32 lines · Marcus Sundberg
Quoted 5 lines · Benjamin LaHaise
What version of
ns83820.care you using? Version 0.17 ofns83820.cmade significant improvements under load. Other possibilities include cabling problems (watch the kernel logs for changes in link state). Try to find out where the packets are getting dropped by looking through/proc/net/snmpand other statistics counters in the kernel.0.17, but some more testing showed that the ns83820 actually works just fine during this test when using just crossover cables and running at gigabit speed. The original testing was done using 100Mbit hubs, so my guess would be that the 83820 chips (and/or driver) doesn't handle collisions too well (which I don't have a problem with, as afaik GE is always switched).
However the DFE-580TX problems remain regardless of using a hubbed or switched network.
(As booth eepro100 and tulip-based cards works fine with the hubs I'm quite certain there's nothing wrong with them.)
//Marcud
Signature · 8 lines
---------------------------------------+-------------------------- Marcus Sundberg <[email protected]> | Firewalls with SIP & NAT Firewall Developer, Ingate Systems AB | http://www.ingate.com/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Quoted 7 lines · On Mon, 3 Jun 2002 [email protected] wrote:
Quoted 4 lines · Marcus Sundberg
running at gigabit speed. The original testing was done using 100Mbit hubs, so my guess would be that the 83820 chips (and/or driver) doesn't handle collisions too well (which I don't have a problem with, as afaik GE is always switched).
There are at least some gigabit ethernet hubs on the market. How badly does it handle collisions?
Huh?
Packet Engines used to sell FDRs -- "Full Duplex Repeaters", but (as the name implies) FDRs were not collision based. I don't know of any Gb Ethernet collision-based repeaters.
Any collision-based operations will take place at 10 or 100Mbps. Anyone still using a repeater obviously doesn't need performance.
...
Quoted 2 lines
Quoted 2 lines
0.17, but some more testing showed that the ns83820 actually works just fine during this test when using just crossover cables and
You might compare against my ns820.c driver. (Note: It was initially released before the ns83820.c driver was written, although the ns83820.c driver was the one put into the kernel.)
Signature · 4 lines
Donald Becker [email protected] Scyld Computing Corporation http://www.scyld.com 410 Severn Ave. Suite 210 Second Generation Beowulf Clusters Annapolis MD 21403 410-990-9993
At 07:56 PM 6/3/02 -0400, you wrote:
Uh..aren't those 32-bit cards? There isn't enough bus bandwidth on a 32bit PCI bus to do gigabit, so its more likely that the cards are overrunning their buffers....
Dennis
Quoted 43 lines
There are at least some gigabit ethernet hubs on the market. How badly does it handle collisions?
NickQuoted 21 lines · Marcus Sundberg
Quoted 5 lines · Benjamin LaHaise
What version of
ns83820.care you using? Version 0.17 ofns83820.cmade significant improvements under load. Other possibilities include cabling problems (watch the kernel logs for changes in link state). Try to find out where the packets are getting dropped by looking through/proc/net/snmpand other statistics counters in the kernel.0.17, but some more testing showed that the ns83820 actually works just fine during this test when using just crossover cables and running at gigabit speed. The original testing was done using 100Mbit hubs, so my guess would be that the 83820 chips (and/or driver) doesn't handle collisions too well (which I don't have a problem with, as afaik GE is always switched).
However the DFE-580TX problems remain regardless of using a hubbed or switched network.
(As booth eepro100 and tulip-based cards works fine with the hubs I'm quite certain there's nothing wrong with them.)
Quoted 11 lines
//Marcud
--
---------------------------------------+--------------------------
Marcus Sundberg <[email protected]> | Firewalls with SIP & NAT
Firewall Developer, Ingate Systems AB | http://www.ingate.com/
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/-
To unsubscribe from this list: send the line "unsubscribe linux-net" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 5 lines · ET Sales
At 07:56 PM 6/3/02 -0400, you wrote:
Uh..aren't those 32-bit cards? There isn't enough bus bandwidth on a 32bit PCI bus to do gigabit, so its more likely that the cards are overrunning their buffers....
32 bit cards are okay for most applications that don't use all bandwidth in both directions. Also, if overruns occur, recent versions of ns83820.c do update the error counts in /proc/net/dev.
-ben
Isn't the DFE-580tx a 4-port 10/100bt NIC? If so, why the talk about Gigabit speed?
Quoted 5 lines · ET Sales
At 07:56 PM 6/3/02 -0400, you wrote:
Uh..aren't those 32-bit cards? There isn't enough bus bandwidth on a 32bit PCI bus to do gigabit, so its more likely that the cards are overrunning their buffers....
Dennis
Signature · 3 lines
Ben Greear <[email protected]> <Ben_Greear AT excite.com> President of Candela Technologies Inc http://www.candelatech.com ScryMUD: http://scry.wanfear.com http://scry.wanfear.com/~greear