Linux HugePages
* What you may already know:
- Larger (usually 2M depending on architecture) memory page size
- Processes share page table entries, less kernel page table management overhead, less TLB misses
- Always resident memory
- On Oracle11g, can be enabled only if AMM is disabled (memory_target set to 0)
- MOS notes 361323.1 (HugePages on Linux: What It Is... and What It Is Not...), 744769.1, 748637.1
- Similar to Solaris ISM (intimate shared memory) in almost every way
Now what you may not know (but may not care either):
* HugePages memory only shows as resident memory on Red Hat 4, not 5
(Actually, it's most likely Linux kernel, not Red Hat version, dependent.) On RHEL 4 server, when
HugePages is used, `top' or `ps' shows that Oracle process's resident memory is only slightly
smaller than virtual memory. But on RHEL 5, resident memory is very much smaller. This, however,
does not change the fact that HugePages memory is guaranteed to be locked in RAM. David Gibson, a
HugePages developer, says in private email "hugepages are always resident, never swapped. This
[RHEL 5 showing non-resident HugePages] must be something changed in the wider MM code".
* vm.nr_hugepages, memlock, and SGA sizing
Process memory lock limit, memlock in /etc/security/limits.conf in KB, can be set to any number
larger than vm.nr_hugepages (in /etc/sysctl.conf, multiplied by Hugepagesize). But it's best to
set vm.nr_hugepages only slightly larger than total SGA's of all instances on the box. (Don't
forget the ASM instance if any.) The memlock setting in limits.conf alone won't actually set
aside memory; it's just a mathematical number to limit the amount of memory locking. So it's OK to
set it to a very big number. But vm.nr_hugepages actually allocates memory. It must not be too
large or your box would run out of memory or can't boot.
If after starting instance HugePages is found to be not used, lower SGA a lot and try again. Then
gradually add SGA back.
* Don't waste memory
If HugePages_Free is not much smaller than HugePages_Total in /proc/meminfo, especially after the
Oracle instance has been running for a while, you're wasting too much memory allocated as HugePages
not being used, because only SysV type shared memory, such as Oracle SGA, can use HugePages. You
can either increase Oracle SGA to close to HugePages_Total (note its unit is Hugepagesize), or
decrease HugePages_Total, preferably by editing /etc/sysctl.conf and running `sysctl -p'. Although
increasing HugePages dynamically may not work because the system free memory may have already
been fragmented, decreasing it always works.
* Checking usage
`cat /proc/meminfo'. Focus on HugePages_* and PageTables. Also, `strace -f -e trace=process
sqlplus / as sysdba' and startup. Look for SHM_HUGETLB in 3rd arg to shmget(). Linux shmat()
doesn't have the option for this flag so tracing listener to follow down to the cloned server
process won't work.[note1] Also, Linux doesn't have -s option for `pmap' as on Solaris to check
page size for the individual mappings inside a process memory space.[note2]
memlock affects shell's `ulimit -l' setting. Make sure your shell has the desired setting before
starting DB instance.
You can check how the numbers HugePages_Free and HugePages_Rsvd change while you startup or
shutdown an instance that uses HugePages (adjust grep pattern as needed):
while true; do
for i in $(grep ^Huge /proc/meminfo | head -3 | awk '{print $2}'); do
echo -n "$i "
done
echo ""
sleep 5
done
The output is like the following (numbers are HugePages_Total, HugePages_Free, HugePages_Rsvd):
512 225 192
512 225 192
512 225 192
512 512 0 <- Instance down. All HugePages freed. (This is the last moment of database shutdown.)
512 512 0
512 371 338 <- Startup. 338 pages free but reserved (i.e. 371-338=33 pages "real" free), 512-371=141 pages used
512 329 296 <- 512-329=183 pages used, up by 183-141=42, reserved pages down by 42, "real" free unchanged
512 227 194 <- 512-227=285 pages used, up by 285-183=102, reserved down by 102 too, "real" free unchanged
It indicates that when the instance is started, HugePages memory pages are immediately reserved.
This is a fast process because there's no write to the pages (remember reserved is just a special
type of free; see ). Then when the pages are written to,
they're taken off of the reserved list and used. This server has 33 "real" free pages wasted. I
could have done better diligence to not assign them to HugePages.
Note that older versions of HugePages code doesn't show reserved pages. On Red Hat Linux, the change is
between RHEL 4 and 5.
* 11g AMM
11g Automatic Memory Management includes PGA into auto management. But PGA can never be allocated
from HugePages memory.[note3] I would set memory_target to 0 to disable AMM and configure HugePages as
usual. HugePages is a far more appealing feature than AMM. If I have to sacrifice one of the two,
I sacrifice AMM. The usage of SGA and PGA is so different they should be separately managed anyway.
To name one issue with AMM, it requires hundreds if not thousands of descriptors for *each* server
process to open *all* the files under /dev/shm, most likely 4 MB each (SGA granule size, _ksmg_granule_size).
See
* In 11g, due to Bug 9251136 "INSTANCE WILL NOT USE HUGEPAGE IF STARTED BY SRVCTL". The root cause
is that process ulimit settings (as in /etc/security/limits.conf) are not used. To confirm, compare
`grep "locked memory" /proc/
/limits' with your setting. You can work around
the bug by setting the limit in either $CRS_HOME/bin/ohasd or /etc/init.d/ohasd, but not /etc/profile
or /etc/init.d/init.ohasd:
# diff /u01/app/11.2.0/grid/bin/ohasd /u01/app/11.2.0/grid/bin/ohasd.bak
5,7d4
< #20100319 YongHuang: Increase process max locked memory to allow HugePages as workaround for Bug 9251136
< ulimit -l 8590000
<
* Further reading
_____________
[note1] On Solaris, you can run `truss -f -p ' and connect to the database through
Oracle Net. The trace will show e.g.
shmat(1979711503, 0x40280000000, 040000) = 0x40280000000
where 040000 is SHM_SHARE_MMU according to /usr/include/sys/shm.h.
[note2] At least not until this patch is in your kernel: That patch
may be in kernel 2.6.29 ().
[note3] For now at least. See Kevin Closson's blog for more:
