1

I am having problems with the network performance between SLES and AIX. I have tested the network performance from

  • AIX 1 Gbit/s to AIX 1 Gbit/s
  • SLES 11 10 Gbit/s to SLES 11 10 Gbit/s
  • AIX 1 Gbit/s to SLES 11 10 Gbit/s and reverse

there are also other machines on the network, so we don't have the full bandwidth, but the network is definitely not flooded by the other machines.

via:

  • netcat
  • scp
  • niping (network performance measurement tool from SAP)

between two AIX machines I am getting "decent" results of about 110 Mbit/s between the two Linux machines I am getting good results of about 2,2 Gbit/s but between Linux and AIX, independent in which direction, I am getting only about 30 Mbit/s, consistently over all the 3 measurement tools. All tested adapters are in the same subnet!

Routing is not the problem. When I am doing traceroute, the nodes are connecting directly to each other without taking a hop over a gateway. There are also no Ierrs/Oerrs according to netstat -i on any machine. Network stability test over ~20 minutes via ping is decent as well.

So personally I would say I can exclude the possibility of a networking problem and can narrow it down to the speed negotiation or buffer size negotiation between AIX and Linux. For your info: All hosts are Logical Partitions (Virtual machines so to speak) on IBM PowerVM. Somebody got an idea what to do?

closed as off-topic by Jeff Schaller, slm Jul 6 '16 at 2:15

This question appears to be off-topic. The users who voted to close gave this specific reason:

  • "Questions describing a problem that can't be reproduced and seemingly went away on its own (or went away when a typo was fixed) are off-topic as they are unlikely to help future readers." – Jeff Schaller, slm
If this question can be reworded to fit the rules in the help center, please edit the question.

  • 1
    110Mbits/s over 1Gbps links between the AIX hosts is terrible - is that a typo and it is in fact 110Mbytes/s? Are they in the same frame? All network going through VIOs? What's their config? Is largesend enabled all the way? ... – Mat Jun 28 '16 at 13:12
  • did you check the duplex settings of the network interfaces on both ends ? If your switch ports set to "auto-negotiate", I have seen more often than not, they set to a lower speed at half duplex on legacy systems, like AIX, and this hampers down the communication speed. – MelBurslan Jun 28 '16 at 14:03
  • Hello experts, thanks for your comments so far. large send and large receive on SLES is turned off, I checked with ethtool. How to check if large send / large receive is on in the AIX LPAR? I tried lsattr -El ent1 but I can'T find the large_send attribute – dafrk Jun 28 '16 at 14:44
  • I suspect there's some confusion over units in the Q; 110 megabytes per second would be a reasonable transfer rate over a 1 gigabit per second line. 30 megabits/s would be ~3.75 megabytes/s. – Jeff Schaller Jun 28 '16 at 15:25
  • nope, sorry, you are wrong. if the transfer rate would be 110 MB/s, there wouldn't be a problem. But we are having problems with our connections between an application server on AIX and the database on SLES, that's why we have poor performance. – dafrk Jun 28 '16 at 20:01
1

try;

  1. swap all cables in the route with new
  2. isolate linux and aix as only two machines on a single switch (reduce the test environment)
  3. cross over patch with static settings (if possible- reducing again)
  4. in case its the switch at fault, test with second switch
  5. get onto the switch itself and check for flapping or errors on the ports
  6. run wireshark on the linux and capture traffic and identify which side the delay is on. or if something else is going on. (if you cant install wireshark locally, then try a promiscuous port and use a third box to monitor)
  7. try with netcat on UDP, as udp has no handshake maybe TCP with its requirements and frame differences is part of the issue. info
  8. move the ports your patched too. check the switch doesnt have any unwanted link aggregation turned on that might play fowl

110mbit/s aix to aix is underwhelming and suggests a possible issue. does aix bench better with anything else?

try transfers to nfs location using mount. and try plain ftp transfers (use a big file, just time it and do the maths to get ball park feel).

note that SCP has encryption overhead and is going to take alot of processing. so steer to un-encrypted transfers for benchmarking/timing.

hope this isolates.

1

The problem lies within the default kernel version of SuSE 11 SP4 being too old. Kernel version from DVD is 3.101.0.63. This kernel version does not support the IBM-proprietary SEA large send and the adjustment of receive and send buffer sizes via ethtool. We upgraded to kernel 3.101.0.100. Then I activated largesend and adjusted size and receive buffer size via ethtool. Now everything works like a charm. I thank you guys for giving me the idea of activating large send because otherwise I wouldn't have found this.

Not the answer you're looking for? Browse other questions tagged or ask your own question.