Date: Fri, 5 Sep 2003 20:31:03 +0300
From: Anatoly Vorobey <mellon@pobox.com>
To: memcached@lists.danga.com
Subject: Re: Memory Management...

On Fri, Sep 05, 2003 at 12:07:48PM -0400, Kyle R. Burton wrote:
> prefixing keys with a container identifier).  We have just begun to
> look at the implementation of the memory management sub-system with
> regards to it's allocation, de-allocation and compaction approaches.
> Is there any documentation or discussion of how this subsystem
> operates? (slabs.c?)

There's no documentation yet, and it's worth mentioning that this
subsystem is the most active area of memcached under development at the 
moment (however, all the changes to it won't modify the way memcached
presents itself towards clients, they're primarily directed at making
memcached use memory more efficiently).

Here's a quick recap of what it does now and what is being worked
on. 

The primary goal of the slabs subsystem in memcached was to eliminate
memory fragmentation issues totally by using fixed-size memory chunks
coming from a few predetermined size classes (early versions of 
memcached relied on malloc()'s handling of fragmentation which proved 
woefully inadequate for our purposes). For instance, suppose
we decide at the outset that the list of possible sizes is: 64 bytes,
128 bytes, 256 bytes, etc. - doubling all the way up to 1Mb. For each
size class in this list (each possible size) we maintain a list of free
chunks of this size. Whenever a request comes for a particular size,
it is rounded up to the closest size class and a free chunk is taken 
from that size class. In the above example, if you request from the 
slabs subsystem 100 bytes of memory, you'll actually get a chunk 128
bytes worth, from the 128-bytes size class. If there are no free chunks
of the needed size at the moment, there are two ways to get one: 1) free
an existing chunk in the same size class, using LRU queues to free the 
least needed objects; 2) get more memory from the system, which we 
currently always do in _slabs_ of 1Mb each; we malloc() a slab, divide 
it to chunks of the needed size, and use them.

The tradeoff is between memory fragmentation and memory utilisation. In 
the scheme we're now using, we have zero fragmentation, but a relatively
high percentage of memory is wasted. The most efficient way to reduce
the waste is to use a list of size classes that closely matches (if 
that's at all possible) common sizes of objects that the clients
of this particular installation of memcached are likely to store.
For example, if your installation is going to store hundreds of                                                                  thousands of objects of the size exactly 120 bytes, you'd be much better
off changing, in the "naive" list of sizes outlined above, the class
of 128 bytes to something a bit higher (because the overhead of 
storing an item, while not large, will push those 120-bytes objects over 
128 bytes of storage internally, and will require using 256 bytes for
each of them in the naive scheme, forcing you to waste almost 50% of
memory). Such tinkering with the list of size classes is not currently
possible with memcached, but enabling it is one of the immediate goals.

Ideally, the slabs subsystem would analyze at runtime the common sizes
of objects that are being requested, and would be able to modify the
list of sizes dynamically to improve memory utilisation. This is not
planned for the immediate future, however. What is planned is the 
ability to reassign slabs to different classes. Here's what this means. 
Currently, the total amount of memory allocated for each size class is
determined by how clients interact with memcached during the initial 
phase of its execution, when it keeps malloc()'ing more slabs and 
dividing them into chunks, until it hits the specified memory limit 
(say, 2Gb, or whatever else was specified). Once it hits the limit, to 
allocate a new chunk it'll always delete an existing chunk of the same 
size (using LRU queues), and will never malloc() or free() any memory 
from/to the system. So if, for example, during those initial few hours 
of memcached's execution your clients mainly wanted to store very small 
items, the bulk of memory allocated will be divided to small-sized 
chunks, and the large size classes will get fewer memory, therefore the 
life-cycle of large objects you'll store in memcached will henceforth 
always be much shorter, with this instance of memcached (their LRU 
queues will be shorter and they'll be pushed out much more often). In 
general, if your system starts producing a different pattern of common 
object sizes, the memcached servers will become less efficient, unless 
you restart them. Slabs reassignment, which is the next feature being 
worked on, will ensure the server's ability to reclaim a slab (1Mb of 
memory) from one size  class and put it into another class size, where 
it's needed more.

-- 
avva