Simpler Navigation for Servers and Operating Systems - Please Update Your Bookmarks
Completed: a much simpler Servers and Operating Systems section of the Community. We combined many of the older boards, so you won't have to click through so many levels to get at the information you need. Check the consolidated boards here as many sub-forums are now single boards.
If you have bookmarked forums or discussion boards in Servers and Operating Systems, we suggest you check and update them as needed.
General
cancel
Showing results for 
Search instead for 
Did you mean: 

mincache=direct & convosync=direct for Oracle

SOLVED
Go to solution
Jean-Luc Oudart
Honored Contributor

mincache=direct & convosync=direct for Oracle

So, I need OnlineJFS product to be able to use these "fine" options when I mount my Oracle data file systems.
Our application would be random & sequential read say 80/20 weight.
1) IS this worth for us to acquire the OnlineJFS product (at what cost)- What is the performance gain (observed not theory please)
2) Do we have to convert existing vxfs file systems and if so what is involved (downtime ?) ?

HPUX 11.0 - Oracle8.1.7.4 64bits - storage on XP128 - with 2 Fibre attachment - RAID1

thanks,
Jean-Luc
fiat lux
10 REPLIES
Stuart Abramson_2
Honored Contributor

Re: mincache=direct & convosync=direct for Oracle

1. Several years ago at another site with an Oracle DB on HP-UX 11.0 we tried convosync and mincache options, not the same as yours, and they didn't do any good at all and we put them back to default.
2. You need Online JFS for these.
3. I don't think you have to convert anything, as long as you have vxfs file systems already.
4. OnLine JFS is worth it's weight in gold because it allows you to change (increase) file system sizes on the fly while the system is running. Have you ever had to take the system down in the middle of the night to increase /usr or /oracle/SID/sapdata25? With OnLine JFS you never have to take the system down or come in off hours to increase file system sizes.
Stefan Farrelly
Honored Contributor
Solution

Re: mincache=direct & convosync=direct for Oracle

1. Cost is expensive for OnlineJFS - in my opinion. Youre talking thousands of $ depending on how large your servers is and how many cpu's/speed. The real benefit of OnlineJFS is you can increase filesystems without having to unmount them - very very handy for production servers.
2. No, you wont have to convert exsiting filesystems. Even if you did the vxupgrade is instantaneous - no downtime needed to upgrade.

Weve tried these options for oracle databases and we found very little performance increase. As well as turning on these mount options you need to tune/modify your oracle database/application to adjust the block size to be a multiple of the filesystem block size, and not all applications do this. If you are prepared to spend quite a bit of time tuning your database to get these options to work fully then the possible performance increase is up to double the current I/O performance, but we didnt think it was worth it.
Im from Palmerston North, New Zealand, but somehow ended up in London...
Sridhar Bhaskarla
Honored Contributor

Re: mincache=direct & convosync=direct for Oracle

Hi Jean,

It is really dependent on how your database is setup. I tried to benchmarking with these options and I got around 5-7% of increase in the throughput.

You do not need to convert the filesystems to make use of OnlineJFS provided if you have Vxfs filesystems. Even you don't need to unmount the filesystems to enable these options. You can use 'remount' option if you want to change these parameters.

Ours is RAID-5 with XP512. We have a large number of RAID groups and have the flexibility of alloting LUNs in the way we want. We do lvm-striping with 128k, 256k and it is giving us significant improvement in the throughput.

-Sri
You may be disappointed if you fail, but you are doomed if you don't try
A. Clay Stephenson
Acclaimed Contributor

Re: mincache=direct & convosync=direct for Oracle

1) OnlineJFS will probably cost you about $3500 - depending upon your platform and discounts, but it's one of those products that every HP-UX box should have whether or not you use it for Oracle - it's just too useful to do without. I have actually seen (and measured) improvements in excess of 20% using the mount options and in some cases as little as 2%. The effects were very noticable under 10.20, less (about half as much) under 11.0, and non-existant under 11.11 where I use cooked files for everything.

If you want to use them, then use convosync=direct,mincache=direct,delaylog,nodatainlog for datafiles and indices and use delaylog,nodatainlog for redo and archive logs.

The main advantage of bypassing the buffer cache was that by limiting the size of the buffer cache, you could increase the size of the SGA but now that systems have very large amounts of memory, I find that cooked files win on 11.11 everytime.

2) No, but you might want to convert your existing Jvxfs filesystems to Version 4 - which can be done "on the fly" and takes a few minutes to complete.
If it ain't broke, I can fix that.
Jean-Luc Oudart
Honored Contributor

Re: mincache=direct & convosync=direct for Oracle

Well,

OnlineJFS does not seem to be much help for me, if I gain *only* 10% !
I understand that the 1st advantage is to increase fs size on the fly, but we are not 24x7 shop and id does not happen that often.
So much for an *expensive* product.

Should I add that AIX "offer" the increase size on the fly.
I think HP should have it as standard feature..

Clay, is vxfs version 4 any good to me re performance ?

thanks,
Jean-Luc
fiat lux
A. Clay Stephenson
Acclaimed Contributor

Re: mincache=direct & convosync=direct for Oracle

Version 4 layout typically offers a few per cent (< 5%) performance increase by more aggressive i/o clustering. More than anything else, OnlineJFS pays for itself in the ease with which filesystems can be expanded (or shrunk) "on the fly". If you combine OnlineJFS with MirrorDiskUX, you can really begin to approach zero downtimes.
If it ain't broke, I can fix that.
Sridhar Bhaskarla
Honored Contributor

Re: mincache=direct & convosync=direct for Oracle

Hi Jean,

If you do striping, then Version 4 JFS may help you quite a bit. It offers tuning options like read_pref_io, write_pref_io that can help you customize your environment for optimal performance.

There are other options like discoverd_direct_io etc., that may be used to improve the performance.


-Sri
You may be disappointed if you fail, but you are doomed if you don't try
Yogeeraj_1
Honored Contributor

Re: mincache=direct & convosync=direct for Oracle

hi jean-luc,

to add to the excellent replies above, i would also recommend reading the following article from Oracle.

Best Regards
Yogeeraj
No person was ever honoured for what he received. Honour has been the reward for what he gave (clavin coolidge)
Jean-Luc Oudart
Honored Contributor

Re: mincache=direct & convosync=direct for Oracle

Thanks for the update

Yogeeraj, I already have the document but as I'm connected to a SAN (XP128), this does not apply really. Writes are done to the cache (therefore the more cache the better) and Reads are from the cache and I suppose it depends on how many spindles you have behind ,... I'm not to sure about how XP128 react with hot spots.

As far as I know the disks used for this server are RAID1+ , it's HP own version for RAID10 (RAID 0+1), eventhough I have not found out what is the stripe size. And if I know it, should I change some of the OS parameters and/or some of the Oracle database parameters

Jean-Luc
fiat lux
Jim Carter_1
Advisor

Re: mincache=direct & convosync=direct for Oracle

Jean-Luc, I am concerned about your last statement regarding "cache". Please don't confuse the cache in the XP128 with the system buffer cache in RAM, nor with the cache in the Oracle SGA.

An advantage of mincache=direct and convosync=direct that hasn't been mentioned is that this setting allows you to isolate your Oracle I/O from your routine system I/O. You can then tune your dbcmax and dbcmin to much more appropriate settings because they are not limiting your actual application cache.

The cache in the XP128 is not tunable by you from a UNIX point of view (I realize that you can tune watermarks within the frame, but that's a different topic). As far as the system is concerned, when either the system buffer cache or the SGA makes a read or write request, the system doesn't really care if it comes from the XP's cache or has to actually come from a spindle.

Someone else has written that they only got 7% increase in performance, but they also wrote that they did not complete the process to fully implement the changes. You should also retune your Oracle engines to take advantage of the new settings. I've had customers tell me they could achieve 15% based on their particular configuration.

Another thought, depending on your particular database, you might not change the performance of the reads or writes, but you might improve the performance of the routine system functions and the applications by isolating the two different types of I/O (oracle or UNIX). Just a thought that I haven't had time to prove or disprove yet.

Just because I'm paranoid doesn't mean they're not out to get me