DS10. EVA 5000 , Slow

[shashi kanth]

I have 2 DS10 boxes, both are having 5.1B with BB27 patch-kit, i have clustered those boxes with TruCluster.

(1) I have assigned 4 LUNs from EVA5K. When i am creating a cluster, when it is copying /,/usr and /var to the cluster /, cluster /usr, cluster /var directories, the copy process taken 2-3 hours

(2) After i finished the creation of a cluster, botted the first member, added second box, and also booted the second member. Everything is fine.

(3) After my cluster came-up, sometimes, the simple commands like ls,mkdir are hangs. These commands takes lot of time to show the output.

(4) But if any point of time, if i boot a box from local disk, the box behaves fast. But if the box boots from a LUN, the box behaves everything slow.

(5) There is no problem with HBA, and there is no problem with Fabric wire. I guess the problem might be with the LUN.

(6) But in my CV EVA, the operational state of that LUN shows "good".

Please any one tell me, how to troubleshoot this issue ?

[Vladimir Fabecic]

Please provide more informations.
What patch kit and additional patches did you install (what is BB27 patch-kit)?
Which FC switches are in fabric? Firmware revision of switches?
Did you see any strange event on FC switch or EVA?
What is used for cluster interconnect, memory channel or ethernet?
Which FC HBA?

[shashi kanth]

Patches installed on the system came from following software kits:
------------------------------------------------------------------

- T64V51BB27AS0006-20061208 OSF540
- T64V51BB27AS0006-20061208 TCR540


I am using Ethernet (cross-over-cable) as cluster interconnect.

#hwmgr -show fibre -ada

ADAPTER LINK LINK FABRIC SCSI CARD
HWID: NAME STATE TYPE STATE BUS MODEL
--------------------------------------------------------------------------------
120: emx0 up point-to-point attached scsi4 FCA-2354

Revisions: driver 2.25 firmware 3.81A4
FC Address: 0x30300
TARGET: -1
WWPN/WWNN: 1000-0000-c930-7da5 2000-0000-c930-7da5


I am not getting any strange errors on EVA/Switch.

[Hein van den Heuvel]

You only talk about the IO setup.
But what about the Cluster Communication?
Since you don't mention it I guess no memory channel, but ethernet.
Fast (GB)? Jumbo enabled? Dedicated? Direct or through infrastructure?

When a new file is created the system which 'owns' the disk (verify with cfsmgr) has to be communicated with. So a poor network can slow down disk IO.

On the disk side you may want to check whether the FC switch is a supported one.

fwiw,
Hein.

[Vladimir Fabecic]

Could be a problem with port of FC switch.
Try to connect FC HBA to another port on FC switch (or to another FC switch).
Once I had strange problem with VMS cluster and robot tape library, and the reason was FC switch.
Problem is probably caused by something between alpha HBA and EVA, i.e. FC cable, GBIC, FC switch or even FC HBA.
What FC switch you have and what is its firmware revision?

[Rob Leadbeater]

Hi,

> - T64V51BB27AS0006-20061208 OSF540
> - T64V51BB27AS0006-20061208 TCR540

OK. So you're using the latest patch kit.


> I am using Ethernet (cross-over-cable) as
> cluster interconnect.

I'm not sure this is supported, but it should work. Can you check whether there's a duplex mismatch on the network cards ?

> 120: emx0 up point-to-point attached scsi4 FCA-2354

This appears to show that you're connected straight to the EVA somehow, rather than through a switch... Are you sure things are cabled properly ?

> Revisions: driver 2.25 firmware 3.81A4

You're running rather old firmware on the FCA-2354. The latest is 3.93a0.

As a first fix, I'd make sure that you install the latest AlphaServer firmware.

http://h18002.www1.hp.com/alphaserver/firmware/index.html

Hope this helps,

Regards,

Rob

[shashi kanth]

---------------------------------------------
I'm not sure this is supported, but it should work. Can you check whether there's a duplex mismatch on the network cards ?
----------------------------------------------

Please can you tell me, how can i check this ?

[Hein van den Heuvel]

"Ethernet" huh?

Care to elaborate? Interface used, Speed.

Maybe post some information from:
# clu_get_info
# hwmgr get attribute -category network
# cfsmgr

Check out: man lan_config

Check out /var/adm/message for the bootrecords indicating the speed/duplux selection.

Or brute force it... sit at console, remove clsuer interconnect cable for a few seconds, and plug back in. Watch the messages closely.

A cluster interconnect deserves (requires IMHO) a gigabit ethernet.

Those tend to not need a cross-over cable.
SO I guess it is runnimg Twisted pair 100mb over a DE600?

Tru64 clusters were really design for using a Memory Channel (or Inifiniband or Quadrix or such) as interconnect: Very low latency, Decent throughput.
100mb Ethernet does not come close to that.

In other words... you may be getting what you paid for!

hth,
Hein.

[Rob Leadbeater]

That depends what NICs you're using...

I can't recall what interfaces are in a DS10 as standard, however I think they're probably Tulip based. From the SRM (P00>>>) prompt, take a look at the EG*0_MODE variables.

P00>>> show eg*mode

Cheers,

Rob

[Rob Leadbeater]

Doh. Sorry that should say EW not EG.

Cheers,
Rob