cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pavel Yaskevich (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-6694) Slightly More Off-Heap Memtables
Date Wed, 09 Apr 2014 07:32:20 GMT


Pavel Yaskevich commented on CASSANDRA-6694:

bq. I'm saying performance critical code is impacted when you have virtual method calls that
cannot be optimised by the VM (i.e. those with multiple implementations). I meant CASSANDRA-6553
and CASSANDRA-6934

Which means that if we actually optimize AbstractType and derivatives to work directly with
underlying bytes whole problem could be resolved? That's why I want to understand why we can't
have a simple implementation of the cell which has one buffer + metadata about component sizes
(which could also be encoded) instead of having buffer per component in the name (if composite)
+ buffer for value + long timestamp? Maybe it would be easier to offload all of the work to
AbstractType instead of trying to optimize on the Cell level?

I went through JVM instruction set doc (specifically
and those
methods are not that different and both have to do lookup in the constant_pool of that class
so I'm wondering if it's virtual calls that create a problem or it's something else masked
by that...

It also looks like if we use static Impl scheme (like in the patch set) would execute the
same amount of instructions because compiler emits *aload_0* (this) in both cases before would
it be invoke\{special, virtual\} or invokestatic, and more instructions in static Impl form
if we use something else instead of "this". Generally when callers use methods from super
class or interface (as it is right now for e.g. Cell.dataSize()) compiler would emit *aload_0,
invokevirtual #offset* directly to the Cell method, where with static Impl it has to that
multiple times *aload_0, invokevirtual #offset* (to the method in DeleteCell.dataSize() and
then internally *aload_0, invokestatic #offset* (to the DeletedCell.Impl.dataSize()) which
means longer constant_pool walk.

bq. Then what exactly do we win? We still have to have two hierarchies and the same modularization.
Also the potential ease of optimizations for comparison disappear, and we still have increased
indirection and virtual method call costs. If this is the suggestion, I am very -1, as the
payoff is very small, the work nontrivial and the negatives substantial.

The wins are, primarily, less object overhead (ultimate goal of all this) and maintainability
of the code. We basically have Cell based on type - expired, deleted, counter, client (the
last one being used mostly by Thrift) as it is right now, so no Buffered* or Native* plus
allocators of 3 types (maybe we actually don't need one which allocates DirectBuffer but can
just go with JNA backed one) which allocate raw bytes. Cell reconcile, equals, dataSize and
other methods become straight-forward. Also, as we consider Composite as a complete entity,
storing components as contiguous blocks would reduce container overhead + speeds up comparisons
by exploiting spatial locality. 

[~jbellis] mentioned this "My preferred solution would be, stop extracting the name so often
by itself. Spot checking the code, it seems we usually do this just to "simplify" a comparison,
so this could in principle just be done with the Cell object rather than just the name." I
think that would would further benefit the approach that I'm describing.

> Slightly More Off-Heap Memtables
> --------------------------------
>                 Key: CASSANDRA-6694
>                 URL:
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Benedict
>            Assignee: Benedict
>              Labels: performance
>             Fix For: 2.1 beta2
> The Off Heap memtables introduced in CASSANDRA-6689 don't go far enough, as the on-heap
overhead is still very large. It should not be tremendously difficult to extend these changes
so that we allocate entire Cells off-heap, instead of multiple BBs per Cell (with all their
associated overhead).
> The goal (if possible) is to reach an overhead of 16-bytes per Cell (plus 4-6 bytes per
cell on average for the btree overhead, for a total overhead of around 20-22 bytes). This
translates to 8-byte object overhead, 4-byte address (we will do alignment tricks like the
VM to allow us to address a reasonably large memory space, although this trick is unlikely
to last us forever, at which point we will have to bite the bullet and accept a 24-byte per
cell overhead), and 4-byte object reference for maintaining our internal list of allocations,
which is unfortunately necessary since we cannot safely (and cheaply) walk the object graph
we allocate otherwise, which is necessary for (allocation-) compaction and pointer rewriting.
> The ugliest thing here is going to be implementing the various CellName instances so
that they may be backed by native memory OR heap memory.

This message was sent by Atlassian JIRA

View raw message