Seconds after uploading my last message, patch-1.3.99.gz comes in.
What do I find, but you've fixed this problem completely!
I'd heard that posting to the net was a good way to get a problem
sorted, but I didn't realise it would be that fast! :-)
I am surprised this went unnoticed for so long. But there you go. It
is fixed now -- this was my major concern for the 2.0 kernel.
Here's some new figures. The network is completely unloaded, but the
local SGI is a bit busy at the moment:
`tcpspray -n 1000' gives:
1.2.13 -> 900k/s (to SGI)
-> 850k/s (to Linux 1.2.13)
1.3.81 -> 700k/s (to SGI)
-> 694k/s (to Linux 1.3.99)
-> 660k/s (to SGI which is busy rendering...)
1.3.99 -> 950k/s (to Linux 1.3.81)
-> 864k/s (to SGI which is busy rendering...)
`etherspray' (my raw packet transmission program) gives:
1.2.13 -> 1197k/s
1.3.81 -> 745k/s
1.3.99 -> 1200k/s
Some additional figures, with 1.3.81 reading from 1.3.99:
cat+smbfs -> 320k/s
smbclient / smbd -> 602k/s
ftp / ftpd -> 717k/s
rsh / cat+rshd -> 674k/s
cat+nfs / nfsd -> 178k/s (8764k/s once it's in the local cache :-)
The two Linux machines involved in this latest round of tests have the
identical motherboards, CPUs and ethernet cards, and much the same other
irrelevant bits. The CPUs are AMD 486DX4/100MHz chips. The NE2k clones
are Trust Computer Products, with "Ethernet Accelerator" chips on board.
They're the fastest ISA NE2k cards of all of the ones I've tested in the
office.
It is nice to see that TCP is now faster than in the 1.2 kernels. After
all the effort that's gone into the TCP code I was wondering if I'd ever
see the benefits. Of course, from these tests it's not possible to say
that TCP transmit time has improved: it may simply be that 1.3.81 is
sending ACKs more smoothly than 1.2.13.
Anyway, congratulations to Paul and everyone else involved. Even with
the NE2k "nothing special" design, Linux is now managing good
throughput again. Hooray!
Cheers,
-- Jamie Lokier