TCP throughput problems

From: system administration account <sysadmin_at_astro.su.se>
Date: Wed, 25 Oct 2000 17:29:34 +0200

Greetings,

I am struggling with the following problem.

Setup: 1 AlphaStation 500 with a DEC 21040-based NIC (10 Mb/s only), attached
to a 10/100 Mb/s switch, either directly or through a 10Mb/s hub; and another
host (e.g., Intel 815E chipset with OpenBSD-current; but the problem is also
seen with other platforms) with a 100Mb/s full-duplex connection. The
AlphaStation is running 4.0E PK4. Similar behaviour is also observed with
a PWS 433au running 4.0D PK6, when forced into 10 Mb/s HD mode.

I am trying to transfer large amounts of data (using rsync 2.4.3 over either
ssh or rsh) from the Digital Unix machine to the other host. The transfer
starts OK, but soon stalls. tcpdump reveals that:

-- the alpha sends pairs of data packets back to back;
-- the first packet makes it to the other host, which immediately ACKs it;
-- the alpha never receives the ACK, the remote never receives the second
packet;
-- the alpha reacts to the lack of ACKs by the expected retransmit and
back-off strategies, but still sends pairs of packets back-to-back. After
a few attempts, it declares the connection broken.

When a Linux (2.2.17 with the via-rhine 1.08 driver) host is substituted
for the alpha (attached to the same 10Mb/s hub), some ACKs are also dropped
but generally don't kill the outbound data packets; and enough ACKs make it
through to keep the pipe from stalling.

I've tried the following:
-- decrease the ipmtu on the alpha. This helps in reducing the error count
reported by netstat, but doesn't prevent the stalling.
-- use rsh instead of ssh; same behaviour.
-- transfer data towards another 10Mb/s host; this works, but I can't run
all of my hosts at 10Mb/s.
-- add some incantations to /etc/iptos (after noting that the Linux box uses
tos=8 (high-throughput) while Digital Unix doesn't). Unfortunately, neither
rshd nor my copy of sshd seem to pay any attention. (The man page doesn't
promise that /etc/iptos will work for anything other than telnet or ftp,
so this isn't entirely surprising.)

My question is: how can I tune DU 4.0D/E to get reasonable throughput in this
kind of situation? (Alternatively, where can I find a 100Mb/s NIC that is
supported under 4.0E? I know about the DE600-AA, but that requires 4.0F with
add-ons, and I'm not keen to upgrade the OS just yet.)

Thanks in advance for any replies; I'll summarize.
Received on Wed Oct 25 2000 - 15:30:38 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:41 NZDT