[plug] file transfer performance linux/windows and filezilla

[Cameron Patrick]

I'm afraid I can't really help you much, but I have some pointless
numbers for comparison...

Denis Brown wrote:

| b)   Transferred an .iso file 650MBytes, same one I used earlier.

Okay, I'm using a 670mb woody iso for my tests, on two ~2GHz Athlon
machines with unremarkable network cards over a generic 100mbit switch.

One machine to another via NFS - 59 seconds, 11.2 mb/s
SSH with default settings - also 59 seconds
FTP - 77 seconds(!), 8.6 mb/s
SSH to localhost - 1m 51s, 6 mb/s

I was surprised at the speed of FTP - while this test is statistically
meaningless it does seem to support Shayne's claim of FTP being
slower.  (Psst, Shayne - the guild e-mail server seems to be running
on its own Murdoch time zone again, your last email was dates tomorrow
morning :-P)

Anyway, the point is that the transfer times you listed below sound
entirely normal.  I presume that they're between the Linux machines,
and not the NT ones that you're complaining about being slow?

| Correct me if I'm wrong but the above would seem to vindicate the network.
| So begs the question why ssh (or Debian's implementaion of ssh???) is so
| slow.

Erm, your figures suggest that it's /not/ slow though?  (i.e. achieving
close to maximal speed on a 100 mbit network.)  Unless you mean slow as
in CPU-hogging, in which case you're right: SSH encryption /does/ have
a massive CPU penalty, and there's no way of avoiding this (besides
buying a faster CPU or using a different encryption method).  This is
why transfers to localhost are slow.  Though you imply later that you
already realise this?

| Just to recap...  several Windows NT4 machines (running FileZilla) are
| equally lethargic for transfers when connected to several Linux (Debian)
| machines, two of which should not lack for horsepower.   So it would seem
| not to be a simple case of one client, or server, being ill or
| ill-configured.

Perhaps the WinNT machines are underpowered, CPU-wise?  Or the network
cards or cables are dodgy?

| [suggestions about using IPs instead of DNS names]
| Based on tonight's experiments I don't think either of these
| straw-clutchings would make much difference.

Neither do I.  Have you tried testing the raw network performance
between the NT and the Linux machines using something like tcpspray?
If you're not getting close to full wire speed on tcpspray, you're
certainly not going to get that speed through SSH.  A reasonable
result would be something like -

$ tcpspray -b 4096 -n 10000 euclid
Transmitted 40960000 bytes in 3.474531 seconds (11512.345 kbytes/s)

Cheers,

Cameron.