Terrible FTP performance 1

[danshelb]

Hi,
I'm having a network performance issue, esp with FTP, on one of my boxes. It's a 7044-270 running 5.2 ML 5. I get something like 40-50 kb/s throughput FTPing from my LAN, whereas all the other boxes in the same datacenter, hung on the same switches, with the same HW and OS level get about 50 times that speed. I've verified port settings on my switch and it matches that of the other servers. I've looked at the NIC and all seems fine and matches the cards in the other "fast" servers. Any thoughts on what I can look at to help determine the root cause of
this?

TIA,
Dan

He who loses honor can lose nothing else.

 

[danshelb]

Never mind, in digging a little deeper I found that despite having the same configuration on the switch as the other boxes, my slow box was being autodetected at 100/half despite the fact the card itself was set at 100/full. Nothing like a little duplex mismatch to cause some performance headaches.

Dan

He who loses honor can lose nothing else.

 

[victorv]

hi,
I should proceed as follows:
- set switch port in Auto (all: speed, mode, ecc.)
- choose a file to download so that the tim to get it is 10
20 seconds
- perform various download from a client of this file,
making the following chenges:

set server nic at

Auto
1000 half and full
100 half and full
10 half

and write down the duration time of the test.

You can find the optimal setting and you may see that downsizing speed, the time to to download decrease: in this case (I belive this is the cause) the patch cable between server and switch has to be changed.


bye

 

[DukeSSD]

victory, NO, not on RS6000 or pSeries machines.

For 1000Mbps adapters on a 1000Mbps network, set the negotiation of adapter to auto or 1000Mbps, if available or auto if not and set the switch to 1000 Mbps if possible or auto if not.

For all 10 and 100 Mbps adapters, or 1000Mbsp adapters being used at 10 or 100 Mbps, set the adapter speed and duplex to the highest the switch / network will support and set / fix the speed and duplex of the switch port to the same.
NEVER try to USE AUTONEGOTIATE on either the switch or the adapter at 10 or 100 Mbps on an RS6000 or pSeries machine, it DOES NOT WORK reliably - as Dan has found.

 

[danshelb]

Duke, you are correct it does *not* work reliably. Any idea whose fault that is, Cisco or IBM's? I've found a few Wintel servers (mostly Dell using Broadcom NIC drivers) that also struggle with auto neg. with Cisco switches. Never got a good answer from Cisco or Dell as to why, we just made sure to always hardcode settings on both sides.

Dan

He who loses honor can lose nothing else.

 

[victorv]

hi all,
probably I am no been clear about the target of the tests:
what I suggested, was intended to discovery
the cause of the problem, no to leave at 10mb a connection
between a Cisco and a pSeries.
The data I wrote, ware only example to make changes and
write down results.

When at the end of the tests you determine certain
situations, you can investigate in a more restricted
field, resolve the problem, an then restore the connection
at the nominal speed.

Many years ago we made a series of changes (new devices) on a network,servers,switch, ecc. with the result of a terrible
lost of performances. Studing, for steps, redoing some connections, we discovered that a patch cable from a pSeries 7043-150 and a switch was no good to substain
a change from 10mb/s to 100mb/s.

At another customer doing these crossover tests, we discovered that the switch was no good for IBM pSeries H70
in HACMP CLuster: then calling the Cisco technician, he upgrade fw of switch, and all was gone OK.

bye

 

[DukeSSD]

The reason IBM give for autonegotiate not being reliable is that although the ethernet protocol standards are very strict and reliable, the autonegotiate standards are very loose and not at all reliable.

Which does kind of make sense - you connect your adapter to a switch, they both detect a link and start trying to autonegotiate, so each starts firing off packets at all sorts of speeds and duplex settings trying to talk to the other, there is a relatively poor chance they will be trying the same speed and duplex at the same time to establish the link, so one gives up and having had a few good packets at, say, 100FD it picks 100FD, At the time though this port wasn’t sending 100FD packets so the other port didn't get any good packets at 100FD and had stopped trying to autonegotiate by the time the first port got fixed at 100FD, but it did get some good 100HD, or 10xx packets, so it goes for 100HD or 10xx. The result - a pants connection and awful throughput.

The only floor in that argument is that just about every other computer on the planet has no trouble getting autonegotiate to work every time - apart from AS400 / iSeries, of course.....

victory, sorry if my reply seemed a bit strong and I appreciate your reasons, but regardless of your experience with AIX, or any other machines, it is very clear cut with RS6k / pSeries / AIX machines as I stated in my reply.