CLoverleaf and I/O to disk

we have cloverleaf 5.8.5 on all VM’s………

All the storage is on SAN.

Our IO is pretty intensive.

We do about 1/3rd of the volume ya’ll do per day

Hi,

I don’t know if it will be useful for you but few weeks ago we have encountered an issue on a 5.8.5 on RHEL about the max “open files” of the system.

A search on lots of files in Global monitor (executed by the hciss) has opened more than 1000 files in the same time (between 2 and 4 files for each real SMAT file) and the system limit was 1024…  it was not a good thing for Cloverleaf !

You can see the system limit with the command : ulimit -a

And the list of opened files with : lsof

Hope it helps

Aurelien

open files are not a limitation; but the latency on our SAN is.

When you do ‘top’……… cpu utilization will come across as a line like this:

Cpu(s): 36.4%us, 42.3%sy,  0.0%ni,  1.7%id, 16.1%wa,  0.0%hi,  3.6%si,  0.0%st

the 16.1%wa is IOWait……. how much time the processor is waiting on on disk.  That, at this point, is our bottleneck, SAN.

open files are not a limitation; but the latency on our SAN is.

When you do ‘top’……… cpu utilization will come across as a line like this:

Cpu(s): 36.4%us, 42.3%sy,

As for the open files, here’s is the related insturctions form the install doc.

System Parameter Settings

You must complete the following steps before installing on your machine:

1 Log in as the root user.

2 Check your current kernel parameters using the following:

>cat /proc/sys/kernel/shmmax

>cat /proc/sys/kernel/sem

>cat /proc/sys/fs/file-max

3 Backup /etc/sysctl.conf and apply the following settings:

Note: There is no need to apply these settings if your current settings are greater than these.

>echo “kernel.shmmax=2147483648” >> /etc/sysctl.conf

>echo “kernel.sem=250 32000 100 1024” >> /etc/sysctl.conf

>echo “fs.file-max=65536” >> /etc/sysctl.conf

4 Reboot your machine or execute the following command to make the parameters effective:

>sysctl -p

Note: You might need to schedule downtime with your user community before rebooting.

You will need to increase the numbers as needed by your environment.

As a follow up….. the things we’ve done have dramatically dropped IO Wait %……..

What we have, is a /sites and /data hierarchy.  Before, we had the entire $HCISITEDIR on /sites and just storing archived SMAT files on data.

What we’ve done now:

/sites still has $HCISITEDIR execpt the $HCISITEDIR/exec/processes tree.

That’s on /data now

So your primary IO, are log files, smat, etc on one disk……. and the raima/recorvery|error|icl database on /sites

So now our IO is split between two disks.

Each disk now has a controller.  In VMWare, there is a virtual controller for /sites, another for /data and a third for everything else.

This seems to have cut our IO wait by 2/3rds.  We were in the 20-40% range on IO Wait %.  Now, most of the time, we’re less than 10%.

VMWare has the same “bottlenecks” that a real controller has….. so this would  work on actual hardware.

Robert:

We have a similar load like you described with 6 million messages a day and SMAT turned on for about everything.

I had concerns about SAN performance because when we were on dedicated disks we had to get creative with dividing the disk and managing what went where physically to keep up.

It might be the actual SAN hardware can vary in design and even utilize SSD drives which can be much faster.

I’m not sure if that is the case here but our SAN performance has been quite delightful for our cloverleaf servers much to my surprise.

The painful part of our SANs have been that they aren’t reliable and go down hard and abruptly.

I was talking to our sys admins about your post and wondering why would we show a meager I/O wait of about 3% on average for similar I/O demands to yours.

Then I was informed by the admins some of our other servers using the same SAN have very bad I/O perfomance, which led our discusion as to why the difference.

One of the observed circumstances with the servers that had bad SAN performance was that they were being mirrorred to another SAN via a low band witdth network capability like to the disaster recovery site.

Investigating this possibility might uncover a possibility not being considered for your situation.

I aslo taked to our admins about dedicated LUN and we are not on a dedicated LUN at this time.

Since we’re on VMWare…… we’ve had to set SYNCFILES=1 in the rdm.ini, which dramatically slowed Cloverleaf (and increased IO) than when it wasn’t on.  However, even with this handicap – Cloverleaf is orders of mangitude faster than the hardware we came from.

We do have dedicated LUNs… and our SAN has SSD’s — however, the IO demands of Cloverleaf aren’t near as great as some of the other apps we have, so Cloverleaf hardly gets to flash.

Greetings,

As you asked for any input here is our situation which is serving us fine

so far regarding SMAT files etc.  We use SAN disk pools.

Running:  5.8.5.0 on AIX 6.1 TL 7 Virtualized

              ulimit -a

              time(seconds)        unlimited

              file(blocks)         unlimited

              data(kbytes)         unlimited

              stack(kbytes)        2097152

              memory(kbytes)       2097152

              coredump(blocks)     2097151

              nofiles(descriptors) 10000

              threads(per process) unlimited

              processes(per user)  unlimited

Sites: 9 all reside on dedicated filesystem allocations (logicial volumes)

Message volume monthly (approx): 12 million

Also run Global Monitor 6.0.1 which opens files as well for SMAT indexing

operations.  This adds to the open files numbers – no hard values at

this time – watching now.

Our history has been that keeping the sites on their own dedicated filesystem (logical volume) keeps the Disk I/O manageable.

When we packed all of our sites on ONE volume then we ran into problems.

We do hourly SMAT archiving using our own scripts rather than vendor

supplied versions.  And keep SMAT files only for 14 days.

We use Tivoli software to archive SMAT files for longer retention.

Then of course work at keeping debug statements out of our TCL,Xlate

inline code to cut DISK I/O in the process logs.

My apologies for not being hardware specific or using the Operating System lingo – more of an application type person than a hardware geek.

Hope this contributes to the dialog and learning here.