Contrasting the SMP/UMA and NUMA architectures
The SMP/UMA architecture
The SMP,
or UMA architecture, simplified
When the PC world first got multiple processors, they were all arranged with
equal access to all of the memory in the system. This is called , or sometimes Uniform Memory Architecture (UMA,
especially in contrast to NUMA). In the past few years this architecture has
been largely phased out between physical socketed processors, but is still alive
and well today within a single processor with multiple cores: all cores have
equal access to the memory bank.
The NUMA architecture
The NUMA
architecture, simplified
The new architecture for multiple processors, starting with and 2processors (we’ll call these “modern PC CPUs”), is a architecture, or more correctly . In this architecture, each processor has a “local” bank of
memory, to which it has much closer (lower latency) access. The whole system may
still operate as one unit, and all memory is basically accessible from
everywhere, but at a potentially higher latency and lower performance.
Fundamentally, some memory locations (“local” ones) are faster, that is, cost
less to access, than other locations (“remote” ones attached to other
processors). For a more detailed discussion of NUMA implementation and its
support in Linux, see .
How Linux handles a NUMA system
Linux automatically understands when it’s running on a NUMA architecture
system and does a few things:
Enumerates the hardware to understand the physical layout.
Divides the processors (not cores) into “nodes”. With modern PC processors,
this means one node per physical processor, regardless of the number of cores
present.
Attaches each memory module in the system to the node for the processor it
is local to.
Collects cost information about inter-node communication (“distance” between
nodes).
You can see how Linux enumerated your system’s NUMA layout using the numactl --hardwarecommand:
# numactl --hardware
available: 2 nodes (0-1)
node 0 size: 32276 MB
node 0 free: 26856 MB
node 1 size: 32320 MB
node 1 free: 26897 MB
node distances:
node 0 1
0: 10 21
1: 21 10
This tells you a few important things:
The number of nodes, and their node numbers — In this case
there are two nodes numbered “0″ and “1″.
The amount of memory available within each node — This
machine has 64GB of memory total, and two physical (quad core) CPUs, so it has
32GB in each node. Note that the sizes aren’t exactly half of 64GB, and aren’t
exactly equal, due to some memory being stolen from each node for whatever
internal purposes the kernel has in mind.
The “distance” between nodes — This is a representation of
the cost of accessing memory located in (for example) Node 0 from Node 1. In
this case, Linux claims a distance of “10″ for local memory and “21″ for
non-local memory.
How NUMA changes things for Linux
Technically, as long as everything runs just fine, there’s no reason that
being UMA or NUMA should change how things work at the OS level.
However, if you’re to get the best possible performance (and indeed in some
cases with extreme performance differences for non-local NUMA access, any
performance at all) some additional work has to be done, directly dealing with
the internals of NUMA. Linux does the following things which might be unexpected
if you think of CPUs and memory as black boxes:
Each process and thread inherits, from its parent, a NUMA policy. The
inherited policy can be modified on a per-thread basis, and it defines the CPUs
and even individual cores the process is allowed to be scheduled on, where it
should be allocated memory from, and how strict to be about those two decisions.
Each thread is initially allocated a “preferred” node to run on. The thread can be run elsewhere (if policy allows), but the scheduler attempts to
ensure that it is always run on the preferred node.
Memory allocated for the process is allocated on a particular node, by
default “current”, which means the same node as the thread is preferred to run
on. On UMA/SMP architectures all memory was treated equally, and had the same
cost, but now the system has to think a bit about where it comes from, because
accessing non-local memory has implications on performance and may cause cache
coherency delays.
Memory allocations made on one node will not be moved to
another node, regardless of system needs. Once memory is allocated on a node, it
will stay there.
The NUMA policy of any process can be changed, with broad-reaching effects,
very simply using as a wrapper for
the program. With a bit of additional work, it can be fine-tuned in detail by
linking in and
writing some code yourself to manage the policy. Some interesting things that
can be done simply with the numactl wrapper are:
Allocate memory with a particular policy:
locally on the “current” node — using --localalloc, and also the
default mode
preferably on a particular node, but elsewhere if necessary — using --preferred=node
always on a particular node or set of nodes — using --membind=nodes
interleaved, that is, spread evenly round-robin across all or a set
of nodes — using --interleaved=all or --interleaved=nodes
Run the program on a particular node or set of nodes, in this case that
means physical CPUs (--cpunodebind=nodes) or on a particular
core or set of cores (--physcpubind=cpus).
What NUMA means for MySQL and InnoDB
InnoDB, and really, nearly all database servers (), present an atypical workload (from the point of view of the
majority of installations) to Linux: a single large multi-threaded process which
consumes nearly all of the system’s memory and should be expected to consume as
much of the rest of the system resources as possible.
In a NUMA-based system, where the memory is divided into multiple nodes, how
the system should handle this is not necessarily straightforward. The default
behavior of the system is to allocate memory in the same node as a thread is
scheduled to run on, and this works well for small amounts of memory, but when
you want to allocate more than half of the system memory it’s no longer
physically possible to even do it in a single NUMA node: In a two-node system,
only 50% of the memory is in each node. Additionally, since many different
queries will be running at the same time, on both processors, neither individual
processor necessarily has preferential access to any particular part of memory
needed by a particular query.
It turns out that this seems to matter in one very important way. Using we can see all of the allocations made by mysqld, and some interesting
information about them. If you look for a really big number in the anon=size, you can pretty easily find the buffer pool (which
will consume more than
51GB of memory for the 48GB that it has been configured to use)
[line-wrapped for clarity]:
2aaaaad3e000 default anon=13240527 dirty=13223315
swapcache=3440324 active=13202235 N0=7865429 N1=5375098
The fields being shown here are:
2aaaaad3e000 — The virtual address of the memory region. Ignore
this other than the fact that it’s a unique ID for this piece of memory.
default — The NUMA policy in use for this region.
anon=number — The number of anonymous pages mapped.
dirty=number — The number of pages that are dirty because
they have been modified. Generally memory allocated only within a single process
is always going to be used, and thus dirty, but if a process forks it may have
many copy-on-write pages mapped that are not dirty.
swapcache=number — The number of pages swapped out but
unmodified since they were swapped out, and thus they are ready to be freed if
needed, but are still in memory at the moment.
active=number — The number of pages on the “active list”;
if this field is shown, some memory is inactive (anon minus active) which means it may be paged out by the swapper soon.
N0=number and N1=number — The number of
pages allocated on Node 0 and Node 1, respectively.
The entire numa_maps can be quickly summarized by the a simple
script numa-maps-summary.pl,
which I’ve written while analyzing this problem:
N0 : 7983584 ( 30.45 GB)
N1 : 5440464 ( 20.75 GB)
active : 13406601 ( 51.14 GB)
anon : 13422697 ( 51.20 GB)
dirty : 13407242 ( 51.14 GB)
mapmax : 977 ( 0.00 GB)
mapped : 1377 ( 0.01 GB)
swapcache : 3619780 ( 13.81 GB)
An couple of interesting and somewhat unexpected things pop out to me:
The sheer imbalance in how much memory is allocated in Node 0 versus Node 1.
This is actually absolutely normal per the default policy. Using the default
NUMA policy, memory was preferentially allocated in Node 0, but Node 1 was used
as a last resort.
The sheer amount of memory allocated in Node 0. This is absolutely
critical — Node 0 is out of free memory! It only contains about 32GB of memory
in total, and it has allocated a single large chunk of more than 30GB to
InnoDB’s buffer pool. A few other smaller allocations to other processes finish
it off, and suddenly it has no memory free, and isn’t even caching anything.
The memory allocated by MySQL looks something like this:
Allocating memory severely imbalanced, preferring Node 0
Due to Node 0 being completely exhausted of free memory, even though the
system has plenty of free memory overall (over 10GB has been used for caches) it
is entirely on Node 1. If any process scheduled on Node 0 needs local
memory for anything, it will cause some of the already-allocated memory to be
swapped out in order to free up some Node 0 pages. Even though there is free
memory on Node 1, the Linux kernel in many circumstances (which admittedly I
don’t totally understand3) prefers to page out Node 0 memory rather
than free some of the cache on Node 1 and use that memory. Of course the paging
is far more expensive than non-local memory access ever would be.
A small change, to big effect
An easy solution to this is to interleave the allocated memory. It is
possible to do this using numactlas described above:
# numactl --interleave all command
We can use this with MySQL by making a , adding the following line (after cmd="$NOHUP_NICENESS"), which prefixes the command to start mysqld with a call tonumactl:
cmd="/usr/bin/numactl --interleave all $cmd"
Now, when MySQL needs memory it will allocate it interleaved across all
nodes, effectively balancing the amount of memory allocated in each node. This
will leave some free memory in each node, allowing the Linux kernel to cache
data on both nodes, thus allowing memory to be easily freed on either node just
by freeing caches (as it’s supposed to work) rather than paging.
Performance regression testing has been done comparing the two scenarios
(default local plus spillover allocation versus interleaved allocation) using
the DBT2 benchmark, and found that performance in the nominal case is identical.
This is expected. The breakthrough comes in that: In all cases where swap use
could be triggered in a repeatable fashion, the system no longer
swaps!
You can now see from the numa_maps that all allocated memory has
been spread evenly across Node 0 and Node 1:
2aaaaad3e000 interleave=0-1 anon=13359067 dirty=13359067
N0=6679535 N1=6679532
And the summary looks like this:
N0 : 6814756 ( 26.00 GB)
N1 : 6816444 ( 26.00 GB)
anon : 13629853 ( 51.99 GB)
dirty : 13629853 ( 51.99 GB)
mapmax : 296 ( 0.00 GB)
mapped : 1384 ( 0.01 GB)
In graphical terms, the allocation of all memory within mysqld has
been made in a balanced way:
Allocating memory balanced (interleaved) across nodes
An aside on zone_reclaim_mode
The zone_reclaim_mode tunable in /proc/sys/vm can be used to fine-tune memory reclamation
policies in a NUMA system. Subject to from the linux-mm mailing list, it doesn’t seem to
help in this case.