Mirrored paging space 1

[sezyigit]

Hi AIX gurus,

We have %100 utilization on local disk drives. there is only OS and paging space on these disks. there are two paging spaces and both are on hdisk0 and mirrored in hdisk1. So bot hdisk1 and hdisk2 are 100% utilized.

I have some questions regarding to descrease disk utilization. I wish you can answer.

1. Is it better to have one paging space big enough instead of two on same disk (16GB one instead of two 8GB ones on same disk)?

2. Also I am confused about mirroring. For paging space it seems to me mirroring may cause utilization. Is mirroring related to high availability? I mean is it a danger if primary paging space goes down or other one?


Best Regards

 

[p5wizard]

If you have a mirrored rootvg, than you have to have mirrored paging space. Makes no sense to have unmirrored paging space, because a disk crash of the disk containing the one paging space copy would still crash the system.

As for sizes, I've read somewhere that it is better to have 8 x 2GB than 2 x 8GB (to do with number of buffers per paging device)


HTH,

p5wizard

 

[sezyigit]

Thanks for the reply.

"Makes no sense to have unmirrored paging space, because a disk crash of the disk containing the one paging space copy would still crash the system."

I understand that there is no need to mirror a paging space. Is it true?

would it be better if we remove paging space copies from hdisk1 and remove second paging space from hdisk0 totally. then create it on hdisk1

Best regards

 

[p5wizard]

No, you misunderstood... perhaps it was a bit unclear?

If rootvg is mirrored, then ALL your paging spaces need to be mirrored (with the copies on different disks of course)in order to survive a disk-crash.



HTH,

p5wizard

 

[bi]

p5wizard is right. You want to mirror your paging space, especially your primary paging space, hd6. If hd6 isn't mirrored but everything else is, if the disk hd6 is on crashes, you won't be able to boot the system. If you have secondary paging space that isn't mirrored, you risk corrupting whatever had paging going on in that paging space if the one disk fails.

I can't tell you how important mirroring is. We had a system that lost one of its mirrored rootvg disks and nothing happened to the system -- it just kept on humming and we could replace the bad disk.

Depending on how much real memory you have, if you are using all of your 16 GB of paging space, you might want to look at tuning some of the vmo (vmtune) values. The defaults are mainly, I think, for systems that don't have a lot of memory. If you don't have a lot of memory, you might want to get more so you aren't paging as much.

As for one big paging space or several little ones, I've heard both sides. I always thought several little ones of equal sizes were the best. Now I work in a place that says one big one is best. I've never seen a good study that supports either way.

 

[kHz]

It is best to spread paging space on as many disks as possible. The 8x2 is better than 2x8 is true. Paging uses round-robin and that is the reason paging space should be equal in size. I would have to dig it out, but an IBM AIX Redbook does state to spread out paging than have it all on one.

 

[costiles]

1) try using the smitty vg to mirror the volume group(s) - you cannot go wrong that way. 2) IBM is telling us to make one big paging space now - for core dumps, etc. And we put them on disks away from the data disks. I learned at an AIX conference to spread them equally over all the disks - that they had to be equal in size. I think that has gone by the wayside. Paging is expensive so we try to tune it out using vmtune or vmi/vmo. IBM delivers a lot of memory configured for caching - we find that we don't need that much. The cost is in I/O - if you don't cache then you use disk i/o - but your users will get more attention from memory if you tune out the paging by tuning for less caching.. We use Universe as our database manager.

 

[kHz]

Wow, another UniVerse user! I last used that DB about 7 years ago. Didn't know there were any other users out there .

 

[plamb]

then why did IBM buy it???

 

[tboege]

Just to add my comment:
It is absolutely a performance killer to place more than one paging space on the very same hdisk! As stated by someone here, the VMM uses a round robin method to access and spread I/Os to the different paging spaces. If you are paging out to many PS on different hdisks, then you're happy as these are concurrently accessible w/o interfering each other. But having 2 or more PS on one hdisk results in head moves and far slower page in/out. This WILL hurt your performance and IBM has release many documents stating that one should never do so!
Have fun!

 

[RodKnowlton]

If you are paging out to many PS on different hdisks, then you're happy

No I'm not.

I'm happy when I'm not paging at all.

Too many people (and vendors) think that paging in AIX is a necessary evil and just live with it. Try this:

- run "vmstat 60" and gather the output long enough to represent all likely workload situations (a week for some, a month for others).
- find the largest value of "avm" in that output
- run "svmon -G" and note the first number, the one at the intersection of "memory" and "size".

If the highest avm value is not near or in excess of the memory size (both values are in 4K pages, btw) and you have paging to paging space (pi and po in vmstat), you need to tune your virtual memory manager. Google for "Tom Farwell" for some excellent articles, or search these forums for "maxperm" to see our many discussions on the subject.

If it is, buy more memory, then proceed to tune.

Accessing memory that's paged out is 100,000 times slower than physical memory. Memory fails less often than disks. It's not a hard purchase to justify (unless your machine is maxed out, in which case you're on your own ).


As always, these are broad statements and your mileage may vary.

Rod Knowlton

IBM Certified Advanced Technical Expert pSeries and AIX 5L
CompTIA Linux+
CompTIA Security+

 

[mrn]

I agree with Rod on this.

I have serveral very busy servers that never page.

I would recommend you have google for sarcheck (If you're not too hot with tuning).

Mike

"A foolproof method for sculpting an elephant: first, get a huge block of marble, then you chip away everything that doesn't look like an elephant."

http://www.airport-parking-site.co.uk/

 

[costiles]

If you are paging - look at tuning maxperm - minperm and maxclient. IBM delivers memory at 80% for caching and 20% for other uses - mainly users. Changing this relationship will help a lot when you are paging excessively!