usbnet ethernet duplex issue?
I'm experimenting with a Netgear FA-120 USB 2.0 to Ethernet device and seeing some strange behavior.
If I run a 100MB transfer (TCP, via nc and dd) over out LAN, with the Netgear on the sending end, I get about 10MB/s, as expected. Receiving, I get ~5MB/s. If I do simultaneous send and receive, the throughput is a few K per second at best.
If I do the same transfers between a pair of isolated laptops, with the Netgear on one end and Intel e100 or e1000 on the other, I see about 500-900K per second in either direction.
There are no errors detected by the usbnet driver and ethtool reports that the device is autonegotiating, full duplex. Setting autoneg off and duplex to half lets the isolated transfers go at wirespeed.
So the question is, what's up with duplex? Everything I can find about the hardware (including the ASIX datasheet) claims it's full-duplex capable but aside from the error counters, it's really behaving like a half-duplex device.
Signature · 1 line
Mathematics is the supreme nostalgia of our time.
Quoted 5 lines · Matt Mackall
So the question is, what's up with duplex? Everything I can find about the hardware (including the ASIX datasheet) claims it's full-duplex capable but aside from the error counters, it's really behaving like a half-duplex device.
Well, USB itself is half duplex, but that wouldn't explain the issue you're seeing. Maybe David Hollis or Phil Chang, who've both worked with the ASIX parts, can cast some light on their quirks. This is on a 2.6 kernel, I hope ... 2.4 kernels have packet queues a bit too high up in the stack to get any real I/O overlap.
I know that when I've done "ttcp" Linux-to-Linux using a net2280 PCI card [1] using CDC Ethernet protocols, I can get a bidirectional transfer rate of around 25 Mbyte/sec (maybe a bit more, I forget; that's untuned), so the bottleneck isn't the network or USB stacks.
- Dave
Quoted 12 lines · Matt Mackall
I'm experimenting with a Netgear FA-120 USB 2.0 to Ethernet device and seeing some strange behavior.
If I run a 100MB transfer (TCP, via nc and dd) over out LAN, with the Netgear on the sending end, I get about 10MB/s, as expected. Receiving, I get ~5MB/s. If I do simultaneous send and receive, the throughput is a few K per second at best.
If I do the same transfers between a pair of isolated laptops, with the Netgear on one end and Intel e100 or e1000 on the other, I see about 500-900K per second in either direction.
Quoted 9 lines
There are no errors detected by the usbnet driver and ethtool reports that the device is autonegotiating, full duplex. Setting autoneg off and duplex to half lets the isolated transfers go at wirespeed.
So the question is, what's up with duplex? Everything I can find about the hardware (including the ASIX datasheet) claims it's full-duplex capable but aside from the error counters, it's really behaving like a half-duplex device.
First off, I hope you are testing with 2.6.12 or so. There are some patches to handle the auto-negotiation that went in 2.6.11 or 2.6.12 and they SHOULD be handling those cases better though it is certainly possible that they don't work as well as they should. In some of my testing, I was finding that I only really got the LPA from autonegotiation on the first connect. Subsequent ones always returned the same values, even if I connected into a 10mb hub or the like. I haven't been able to determine if that is a shortcoming in the device or not.
Signature · 1 line
David Hollis <[email protected]>
Quoted 23 lines · David Hollis
Quoted 12 lines · Matt Mackall
I'm experimenting with a Netgear FA-120 USB 2.0 to Ethernet device and seeing some strange behavior.
If I run a 100MB transfer (TCP, via nc and dd) over out LAN, with the Netgear on the sending end, I get about 10MB/s, as expected. Receiving, I get ~5MB/s. If I do simultaneous send and receive, the throughput is a few K per second at best.
If I do the same transfers between a pair of isolated laptops, with the Netgear on one end and Intel e100 or e1000 on the other, I see about 500-900K per second in either direction.
Quoted 9 lines
There are no errors detected by the usbnet driver and ethtool reports that the device is autonegotiating, full duplex. Setting autoneg off and duplex to half lets the isolated transfers go at wirespeed.
So the question is, what's up with duplex? Everything I can find about the hardware (including the ASIX datasheet) claims it's full-duplex capable but aside from the error counters, it's really behaving like a half-duplex device.
Quoted 9 lines
First off, I hope you are testing with 2.6.12 or so. There are some patches to handle the auto-negotiation that went in 2.6.11 or 2.6.12 and they SHOULD be handling those cases better though it is certainly possible that they don't work as well as they should. In some of my testing, I was finding that I only really got the LPA from autonegotiation on the first connect. Subsequent ones always returned the same values, even if I connected into a 10mb hub or the like. I haven't been able to determine if that is a shortcoming in the device or not.
This appears to be fixed as of 2.6.12.1. My final test before 2.612 was running UDP streams with netcat. It appeared to be dropping just about all outbound packets without throttling the sender and without reporting any errors. Now all seems to be behaving correctly, 11MBps in both directions simultaneously.
Signature · 1 line
Mathematics is the supreme nostalgia of our time.