topic Re: Your opinion on HBA performance data you would like to see in Operating System - HP-UX

Your opinion on HBA performance data you would like to see

marie-noelle jeanson_1 — Thu, 04 Aug 2005 11:39:15 GMT

Hi,

HP is seeking your opinion on what performance data would be most useful to return in tools like sar or kmetrics for Host Bus Adapters (HBA) on HP-UX systems?
- max throughput
- number of IOs per second
- utilization (% theoritical max throughput)
- average IO queueing/response times
- other?

Feel free to add any comments as to why/how you would use these performance data and which perf tool you use the most. Thank you in advance,

Marie-Noelle (points to be assigned!)

Re: Your opinion on HBA performance data you would like to see

Florian Heigl (new acc) — Thu, 04 Aug 2005 11:53:57 GMT

usually I first check if I get at least 30% of the theoretical performance for a raw device read with dd, then for two reads, then I look at i/o-wait with sar -u and avqueue values with sar -d

i.e. like this:
for platterl in /dev/rdsk/c5t0d1 /dev/rdsk/c7t0d0; do timex dd if=${platterl} of=/dev/null bs=1024k count=200 & done

gathering wio should be done with real application load, because it'll always be close to 100% for a dd (dd basically does nothing but wait for data to push to /dev/null, so what else should I expect)
It's much more interesting combined with aqueue/avserv values under regular database load.

sar -d 3 1 during 2GB dd's over two adapters.
18:46:26 device %busy avque r+w/s blks/s avwait avserv
18:46:29 c1t15d0 2.33 0.50 4 24 2.85 14.75
c3t15d0 2.33 0.50 3 21 3.42 12.93
c7t0d0 98.34 0.50 466 238820 4.92 2.11
c5t0d1 98.67 0.50 475 243072 5.00 2.07

If I worked at a research or CGI site, I'd probably go dig into the performance values and get closer to what marketing people call 'line rate', maybe expecting 300-350MB/s read + 350MB/s write at the same time[*], but where I am the performance requirements are more targeted towards database throughput, which is largely achieved by striping and whatever.

[*] everyone who's about to say these numbers are never achieved in real situations, please just don't. a five year old octane2 could handle well above >600MB/s. :)

Re: Your opinion on HBA performance data you would like to see

Steven E. Protter — Thu, 04 Aug 2005 12:05:01 GMT

My input is an extension of glance that specifically provides data on the HBA cards and their performance.

I'd like to see the tools provided for normal disk performance ported to work with data we have to use fcmsutil to get now.

A full toolset is really needed. Everything you list and more.

SEP

Re: Your opinion on HBA performance data you would like to see

Olivier Masse — Thu, 04 Aug 2005 12:12:30 GMT

Personnally I prefer the Mesureware agent over sar, as there are lots of tools available to extract and interpret the data. I think it's included in all the OEs except the foundation OE now, so most of your customers probably have it installed. It might be better to dedicate effort in enhancing and documenting Mesureware metrics rather than sar. Just my two cents.

Olivier

Re: Your opinion on HBA performance data you would like to see

Tim D Fulford — Thu, 04 Aug 2005 15:45:54 GMT

OK

IO/s read
IO/s write
kB/s read
kB/s write
IO read latency
IO write latency
port queue (host side)
% utilisation **
IO/s errors/resends/corrupt/CRC ...

** I would like to see % utilisation, but it would depend upon how it is calculated. We use really quite (probably too small) small IO's, so the HBA is most likely to be limited by the max IO/s which may well only be 10% of the max kB/s limit. So I would like a % util calculated like so (if at all possible)

[total time - ammount of time HBA is idle]/total time

The idle time of the HBA must be time that IO could be placed on the HBA. In the case of IO/s limited systems the HBA will have times when it could theoretically process IO, but is in reality unable to as the OS/card is blocked performing other functions/context switching. I guess that this would mean the %util would need to be calculated from the OS point of view, as that would probably be the limiting factor... but there may be better men/women than me who know more

Regards

Tim

Re: Your opinion on HBA performance data you would like to see

TwoProc — Thu, 04 Aug 2005 16:37:17 GMT

Please get the data and measurement into Glance. It *really* needs updating to cover the new things added to HPUX. I don't think I've noticed anything new in Glance since I first loaded it in 1997, meanwhile HPUX has added lots of options to its O/S.

Re: Your opinion on HBA performance data you would like to see

Tim D Fulford — Thu, 11 Aug 2005 03:19:16 GMT

off track, but Glance should also monitor tape performance too (like it does disks)... or any peripheral IO device/class

Just a suggestion

Regards

Tim

Re: Your opinion on HBA performance data you would like to see

Mahesh Kumar Malik — Thu, 11 Aug 2005 03:29:12 GMT

Hi

I feel error counts in terms of single bit failures if any should also be included

Regards
Mahesh

Re: Your opinion on HBA performance data you would like to see

Cem Tugrul — Thu, 11 Aug 2005 09:43:11 GMT

Glance is my favourite but on the other
hand if we are talking HBA so means there
should be a SAN env and We have to look
SAN side...Then for EVA i also prefer to use
Hp command view EVAperf also gives many
important statistics like;

Arrays
Arrays controllers
Host connections
host port statistics
physical status
port status
.
.

just a suggestion as Tim

Greetings from Turkey,

Re: Your opinion on HBA performance data you would like to see

marie-noelle jeanson_1 — Thu, 11 Aug 2005 09:50:57 GMT

Cem,

HBAs are not only for SAN, there are also WAN adapters (LAN, X.25, ATM, ...). So your reply is not valid for all HBAs.

Marie-Noelle

Re: Your opinion on HBA performance data you would like to see

Devesh Pant_1 — Thu, 11 Aug 2005 11:46:11 GMT

In addition to what you have already mentioned,
I would also like to see the I/O errors with some information to troubleshoot where that error is originating.

thanks
Devesh

Re: Your opinion on HBA performance data you would like to see

generic_1 — Thu, 11 Aug 2005 23:28:03 GMT

Ad a command that shows light dbm if the chipsets support it. This would be great for finding bad equipment.