Recently I've noticed the occasional thread in Oracle newsgroups and lists asking about hugepages support in Linux, including 'best practices' for hugepages configuration. This information is out on the 'world-wide web' in various places; I'd rather put a lot of that information in this article to provide an easier way to get to it. I'll cover what hugepages are, what they do, what they can't do and how best to allocate them for your particular installation. Let's get started.
"Normal" memory pages in Linux are 4 KB in size and are allocated as needed where the memory map will allow so they are likely not contiguous. Hugepages, in comparison, are 2 MB pages locked in memory and are allocated in a contiguous 'chunk'; these are allocated at boot time using a parameter in the /etc/sysctl.conf file named vm.nr_hugepages for RedHat Linux kernel 2.6 and a parameter named vm.hugetlb_pool for RedHat Linux kernel 2.4. You must remember that each page is 2 MB is size, as this affects how many hugepages you need to allocate to cover all of the SGAs of running Oracle databases. Set this too high and your system might not boot; set this too low and you won't get the benefit of using hugepages in one or more Oracle instances. [When use_large_pages is set to ONLY and insufficient hugepages are available the database won't start; changing use_large_pages to TRUE allows the database to start but it won't be using hugepages, which results in more paging activity, lowering performance.] Since hugepages are contiguous it stands to reason that performance will improve since Oracle won't need to access memory segments across the entire range of installed memory. Hugepages also reduce the size of 'page tables' by reducing the overhead for page tables for each connection to the database. Page tables are part of the Linux memory management, created and used to map a virtual address to a physical memory location. For SGAs 100 MB and greater using hugepages is probably a good idea, and that idea gets better as the SGA size increases. Kevin Closson reported back in 2009 on the page table overhead for an Oracle database with an 8000 MB SGA running 500 simultaneous connections. The page table overhead reported was startling -- almost 6.8 GB of memory used for page tables when hugepages were not configured. He ran the same test again, on the same Oracle database, only that test was run after hugepages were allocated and used by the Oracle database. The page table overhead was reduced to slightly less than 23 MB. That is a dramatic change in memory usage, one that can definitely improve performance.
Hugepages are available for IPC (Inter-Process Communication) shared memory; this is the 'standard' shared memory model (starting with UNIX System V) allowing multiple processes to access the same shared memory segment. There is also another form of shared memory segment, the memory-mapped file, and currently such segments are not supported by hugepages. Oracle, on Linux, gives you a choice of using hugepages or memory-mapped files and you implement that choice by selecting to use (or not use) Automatic Memory Management (AMM). Choosing AMM disallows hugepages since AMM uses memory-mapped files located in /dev/shm; you cannot use AMM and hugepages together. For databases with SGAs up to 100 MB using AMM is probably a good idea; a script to determine how many hugepages to allocate won't consider SGAs of less than 100 MB. For Oracle databases with larger SGAs hugepages is, in my opinion, the choice to make as you can significantly reduce shared memory overhead where page tables are concerned.
Calculating how many hugepages to allocate has been a difficult task in the past; Oracle supplies a script that can calculate that value for you provided you have all of the databases that you want to use hugepages running; the script from Oracle won't consider databases you have created that aren't up and available as it uses the ipcs -m command to return the allocated shared memory segments. The script is available from My Oracle Support in Doc ID 401749.1. It generates a list of IPC shared memory segment sizes owned by Oracle and computes the total number of hugepages required. It outputs the result to be used in setting the vm.nr_hugepages parameter in /etc/sysctl.conf. Once that setting is modified the system has to be rebooted before the changes are implemented. Sample output from that script is shown below:
This script is provided by Doc ID 401749.1 from My Oracle Support
(http://support.oracle.com) where it is intended to compute values for
the recommended HugePages/HugeTLB configuration for the current shared
memory segments. Before proceeding with the execution please note following:
* For ASM instance, it needs to configure ASMM instead of AMM.
* The 'pga_aggregate_target' is outside the SGA and
you should accommodate this while calculating SGA size.
* In case you changes the DB SGA size,
as the new SGA will not fit in the previous HugePages configuration,
it had better disable the whole HugePages,
start the DB with new SGA size and run the script again.
And make sure that:
* Oracle Database instance(s) are up and running
* Oracle Database 11g Automatic Memory Management (AMM) is not setup
(See Doc ID 749851.1)
* The shared memory segments can be listed by command:
# ipcs -m
Press Enter to proceed...
Recommended setting: vm.nr_hugepages = 25240
For Linux RedHat kernel 2.6 this is the number of hugepages, not the overall size of the memory allocated for those hugepages. For RedHat kernel 2.4 a different parameter, vm.hugetlb_pool, is set to the actual size that pool needs to be. Fortunately the script knows which kernel is running and reports the correct parameter and value to set.
After you have hugepages allocated you need to tell Oracle to use them exclusively. An init parameter, use_large_pages, should be set to ONLY to keep Oracle from allocating conventionally managed memory segments for the SGA. If your hugepages calculations are correct then all of the databases on your database server will start without error. Allocate too few hugepages and databases with use_large_pages set to ONLY won't start after all of the existing hugepages are used. Setting use_large_pages to TRUE will allow a database to start without using hugepages but that's not what you want to get better performance from Oracle.
Earlier I mentioned that allocating too many hugepages can be worse than allocating too few hugepages. Since each hugepage consumes 2 MB of space multiplying the vm.nr_hugepages setting by 2097152 will show how much memory is occupied by hugepages. It is possible to try to allocate all of the system memory as hugepages resulting in a system that will not boot in multi-user mode. The kernel isn't configured to use hugepages; the more hugepages that are allocated the less conventionally managed memory is available. Take it all away and there is no place for the operating system to start necessary processes after the kernel is loaded at boot time. Using single-user mode will allow the system administrator to modify the offending entry and change it to a more reasonable value so the system can be rebooted. Of course you don't want to end up in that situation; manually calculating such a value can result in errors so use the Oracle-supplied script to calculate the hugepages value you actually need.
Exadata systems run best when hugepages are allocated and used exclusively. The exachk script checks for such settings in each running database on the system and reports those that do not have use_large_pages set to ONLY. It also checks if the currently configured hugepages setting meets or exceeds the estimated number of hugepages required by all running instances on the given database (compute) node. If this check fails it most likely will fail for all available nodes on the system. Failing this check isn't the end of the world but it does mean that the failing databases aren't using memory as efficiently as they could. For each node run the script, make the reported changes then reboot. Before starting any of the databases ensure that all have use_large_pages set to ONLY; start the databases across the cluster using srvctl. The next run ef exachk should report that the hugepages check passed.
RAC databases need to be configured the same way Exadata databases are configured (since Exadata databases are RAC databases). Modify kernel parameters on every database server in the cluster, then reboot. Modify the spfile for each database in the cluster then use srvctl to start those databases cluster-wide. Presuming the hugepages settings are correct across the cluster there should be no error when starting the databases.
Using hugepages for Oracle databases on Linux is, to me, the only way to go if you want to increase database performance. Of course, if you have loads of memory to spare you may not see the need. Most systems I've worked on don't have that luxury so being as 'frugal' with memory as you can is good. Making the most of available resources is key and hugepages on Linux can definitely help with respect to Oracle.
See all articles by David Fitzjarrell