Create/Use hugepage memory on PA-RISC v11.11?

[Jeff Crew]

How do you create HP-UX hugepage memory on a PA-RISC server w/ HP-UX v11.11, so it can be used by an Oracle database?

I recently attended an Oracle conference where creating and using "hugepage" memory structures was discussed. Then the Oracle database would use that "hugepage" memory (structure) for its SGA memory needs.

[smatador]

Hi, Hugepage is for linux is't it?. I never heard this for HPUX.
On linux
cat /proc/meminfo |grep Hugepagesize

[smatador]

Too much for me, but maybe is our pseudo-swap
http://forums11.itrc.hp.com/service/forums/questionanswer.do?threadId=793254

[Jeff Crew]

Most of the chat from the 2 sessions I was in was on Lniux, but they also were talking about AIX and HP-UX Itanium.

One presentor I talked to had a client do "hugepages" or "largepages" on Itanium but hadn't heard of anyone doing it with PA-RISC.

[Don Morris_1]

Yes, these are usually called large pages or super pages in HP-UX.

11.11 already comes with large page support. There are three kernel tunables that directly influence things -- vps_ceiling, vps_chatr_ceiling and vps_pagesize. [Here's the v1 tunable documentation link: http://docs.hp.com/en/TKP-90202/index.html ]

The Oracle application/DB can then use the chatr(1) command to set a hint for page size for Data -- or the kernel will use a heuristic to choose a large page size for the SGA automatically. The heuristic is what vps_pagesize and vps_ceiling control.

Given that vps_ceiling is limited to 64Kb pages -- chatr is the preferred option (chatr +pd L). Which binaries to do this to may be what the conference was about.

[Jeff Crew]

Don - I'll have to look at that doc link. So with my production db SGA of 5809Mb how might that be configured with those kernel params you mentioned?

[Don Morris_1]

vps_chatr_ceiling=419304

[That's the maximum and allows a 4Gb page].

That and the chatr +pd L would be the way I'd go -- if the system has them and supports them (I don't remember off hand which PA supports 4Gb pages vs. 1Gb or if all do, sorry.), you would get 1 4Gb page + 1 1Gb page + 3 256Mb pages + 2 16Mb pages + 2 4Mb pages + 1 1Mb page for you 5809Mb total. That would be the best layout -- and would require pages of that size being available as free memory and the SGA virtual alignment working out (by default on newer releases SysV shmem alignment is done based on the segment size to try to preserve optimum large page sizes -- but I don't think v1 ever got that tweak so you may get a few more 1Gb pages instead of the 4Gb page anyway).

You can check and see what you actually got using pstat(2). The pstat_getprocvm() call returns this as part of the struct pst_vm_status structure it fills in (the pst_vps_pgsizes array -- just walk from 0 to PST_N_PG_SIZES to get the count at each index -- each index is (1 << index) base pages in size [so index 1 is 2 4k pages or an 8kb page, for example].

I'm attaching a sample program to get you started if you need it... but since I don't have an 11.11 machine available, I won't promise it will compile with no effort. It should be pretty close though (just some extra stuff from later releases that I've tried to keep out). You'd want to add a loop like I mentioned above after line 192 or so [so you dump the page size for each successful pstat_getprocvm() call].

[Don Morris_1]

Bleah... you'd think I could look at one window and type in another. Make that "4194304". Don't know why I missed the 4... obviously, that's 4 * 1024 * 1024 if you want to check the math.

[Don Morris_1]

Ah... found an older program. This was written in the 11.0 / 11.11 timeframe so should be good.

I did add "512M", "1G", "4G" and set N_PG_SIZE_NAMES to 20 instead of 17 -- but you may not need to do that.

[Jeff Crew]

Don - I found this informative whitepaper which is helping to tie your notes to someother pts I'm learning.
http://docs.hp.com/en/1218/mem_mgt.html