*BAD* impact of usb on PCI performance

*BAD* impact of usb on PCI performance

[Pavel Machek]

Hi!

I did some testing how does usb hurt rest of system; and it can be
pretty bad.

With Acher's uhci, even ifconfig up of usb-to-usb networking device
[plusb handled by generic usb-to-usb driver; see -ac series].
does 50% slowdown. When fsbr is being used, systems slows down by 350%
(running more than 4 times slower than normally. Ouch).
								Pavel
2.4.5: 300MHz, no usb at all (compiled out)
0.06user 3.25system 3.30 (0m3.308s) elapsed 100.00%CPU
0.06user 3.25system 3.30 (0m3.307s) elapsed 100.00%CPU
0.03user 3.28system 3.32 (0m3.323s) elapsed 99.61%CPU
2.4.5: 300MHz, usb ifconfig down [acher]
0.05user 3.28system 3.35 (0m3.357s) elapsed 99.18%CPU
0.03user 3.29system 3.31 (0m3.319s) elapsed 100.00%CPU
0.04user 3.29system 3.32 (0m3.325s) elapsed 100.00%CPU
0.01user 3.32system 3.32 (0m3.326s) elapsed 100.00%CPU
2.4.5: 300MHz, usb not plugged [JE]
0.06user 3.24system 3.30 (0m3.303s) elapsed 99.90%CPU
0.06user 3.25system 3.30 (0m3.304s) elapsed 100.00%CPU
0.02user 3.30system 3.33 (0m3.336s) elapsed 99.52%CPU
2.4.5: 300Mhz, plusb loaded and UP [acher]
0.03user 4.93system 4.99 (0m4.993s) elapsed 99.33%CPU
0.02user 4.94system 4.99 (0m4.991s) elapsed 99.37%CPU
0.04user 4.92system 4.99 (0m4.990s) elapsed 99.40%CPU
0.06user 4.90system 4.98 (0m4.985s) elapsed 99.49%CPU
0.04user 4.91system 4.98 (0m4.983s) elapsed 99.33%CPU
0.05user 4.91system 4.98 (0m4.985s) elapsed 99.50%CPU
0.06user 4.90system 4.98 (0m4.989s) elapsed 99.42%CPU
0.06user 4.91system 4.99 (0m4.995s) elapsed 99.49%CPU
2.4.5: 300Mhz, plusb plugged in [JE]
0.03user 3.29system 3.35 (0m3.352s) elapsed 99.05%CPU
0.02user 3.29system 3.34 (0m3.349s) elapsed 98.85%CPU
0.01user 3.30system 3.34 (0m3.349s) elapsed 98.84%CPU
2.4.5: 300Mhz, plusb plugged in and UP [JE]
0.04user 3.29system 3.32 (0m3.327s) elapsed 100.00%CPU
0.02user 3.31system 3.35 (0m3.356s) elapsed 99.23%CPU
0.04user 3.28system 3.35 (0m3.353s) elapsed 99.03%CPU
2.4.5: 300Mhz, plusb loaded and UP, fsbr forced ON [acher]
0.06user 14.39system 14.54 (0m14.542s) elapsed 99.37%CPU
0.03user 14.51system 14.62 (0m14.625s) elapsed 99.42%CPU
0.04user 14.63system 14.79 (0m14.798s) elapsed 99.14%CPU
2.4.5: 300Mhz, plusb loaded and UP, fsbr forced ON [JE]
0.04user 14.87system 14.98 (0m14.986s) elapsed 99.49%CPU
0.06user 14.85system 14.99 (0m14.994s) elapsed 99.43%CPU
0.04user 14.85system 14.98 (0m14.984s) elapsed 99.37%CPU

-- 
I'm pavel@ucw.cz. "In my country we have almost anarchy and I don't care."
Panos Katsaloulis describing me w.r.t. patents at discuss@linmodems.org

Re: [linux-usb-devel] *BAD* impact of usb on PCI performance

[David Brownell]

> From: "Georg Acher" <acher@in.tum.de>
> On Sat, Jun 02, 2001 at 01:22:00PM +0200, Pavel Machek wrote:
> > 
> > With Acher's uhci, even ifconfig up of usb-to-usb networking device
> > [plusb handled by generic usb-to-usb driver; see -ac series].
> > does 50% slowdown. When fsbr is being used, systems slows down by 350%
> 
> Hm, the 50% make me curious... have to look what's happening...

For PL-2301/2302 devices, "ifconfig up" is mostly just posting a bulk read.
True with both "usbnet" and its (now obsolete) predecessor drivers "plusb"
and (for different devices) "net1080".

Laplink-style cables can often support another mode (poll via USB "interrupts",
and then issue reads only when data is available) but not every device can work
that way (like, I seem to recall, an iPaq PDA).  And that'd increase the latency
per packet by a couple milliseconds, even when it's possible.


> > (running more than 4 times slower than normally. Ouch).
> 
> Blame Intel. Either low latency or low PCI usage, you can choose...

This problem is UHCI-specific, not USB-generic, yes?
Doesn't happen with OHCI?

- Dave