[vortex] Scyld vs kernel drivers

[Donald Becker]

On Fri, 6 Dec 2002, Richard Fuchs wrote:

> >>>I was wondering about the differences between Donald Becker's driver and
> >>>the driver that comes with the linux kernel,
> > ...
> >>>especially the differences performance-wise.
> > Every modern PCI NIC can get full wire rate performance.
> > The difference is how much system load the NIC imposes.
> > This is impacted by:
...
> >    TCP/UDP/IP checksum calculations (the 905B/905C are excellent)
...
> Perhaps I should have clarified more, I was actually thinking about cpu 
> usage and system load and not so much transfer rate. I'm using two 
> 3c905C cards and I'm looking for ways to cut down the cpu usage. I 
> looked at the driver sources some and read the docs, but it still isn't 
> clear to me how the driver handles for example hardware checksumming 
> (somehow related to zerocopying).

The 3c905B and C have full hardware checksumming, as opposed to the more
common "partial checksum" support.  And the chips do it exactly right.

On receive the chip looks at the type of the incoming packet (IPv4,
IPv4+UPD, IPv4+TCP) and provides the "checksum passed" bits.  The driver
merely needs to enable the feature, and looks at the bits.  If the
checksum doesn't pass, an uncommon error, the driver lets the protocol
level code re-checksum and increments whatever error counters it is
keeping.

Transmit checksumming is much less important.  The data is still hot in
the cache so the software can copy-align and checksum efficiently.  And
the NIC must transfer and interpret the whole packet before stuffing the
checksum into the header, which adds latency.

"Zero-copy" sounds very sexy to work on and has high marking value, but
is applicable on in limited circumstances.  Locking down a single page
on an SMP can be more expensive than just copying in-cache data to
where it needs to be.  So you must be in a situation where
  you have 32KB+ to transmit
  the TCP receiver's window has opened to 32KB+ (!)
  that data is already in lockable memory pages
  the data will not be written during the transmit
  the pages are accessible to the PCI bus

Having a NIC that computes the Tx checksums is a requirement for
"zero-copy" transmit to make any sense, otherwise the CPU has to touch
all of the bytes anyway.  Since you must deal with large blocks, you
really want full TCP segmentation support in the NIC.

> To make a long story short: can I expect a drastic decrease in cpu
> usage if I switched from the kernel driver to your driver?

No.

> Or is there other hardware that does a better job at taking load off
> the cpu?

It depends on what you are doing.  A cluster compute node or application
server node is a much different environment than a workstation.

-- 
Donald Becker				becker@scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Scyld Beowulf cluster system
Annapolis MD 21403			410-990-9993