[infinispan-dev] Adaptive marshaller buffer sizes - ISPN-1102

Thu Jun 9 15:06:35 EDT 2011

Hi guys

This is an excellent and fun discussion - very entertaining read for me.  :-)  So a quick summary based on everyones' ideas:

I think we can't have a one size fits all solution here.  I think simple array copies work well as long as the serialized forms are generally small.  And while I agree with Bela that in some cases (HTTP session replication) it can be hard to determine the size of payloads, in others (Hibernate 2LC) this can be determined with a fair degree of certainty.

Either way, I think this can be a bottleneck (both in terms of memory and CPU performance) if the serialized forms are large (over 100K? That's a guess... ) and buffers are sub-optimally sized.

I think this should be pluggable - I haven't looked at the code paths in detail to see where the impact is, but perhaps different marshaller implementations (maybe all extending a generic JBoss Marshalling based marshaller) with different buffer/arraycopy logic?  So here are the options I see:

1) Simple arraycopy (could be the default)
2) Static buffer size like we have now - but should be configurable in XML
3) Adaptive buffer (the current Netty-like policy Galder has implemented, maybe a separate one with reservoir sampling)
4) Per-Externalizer static buffer size - Externalizer to either provide a deterministic buffer size or a starting buffer size and growth factor.  

Option (4) would clearly be an "advanced option", reserved for use by very experienced developers who want to squeeze every drop of performance, and have intimate knowledge of their object graphs and know what this demands of their system in terms of serialization.

But further, we should also have some logging in the marshaller - probably TRACE level, maybe JMX, disabled by default - to monitor samples and gather statistics on inefficiently configured buffer sizes and policies, perhaps even log marshalled types and resulting sizes.  This could be run during a stress test on a staging environment to help determine how to tune marshalling based on the policies above.

WDYT?  I think the benefit of making this pluggable is that (a) it can be done piece-meal - one policy at a time and (b) each one is easier to unit test, so fewer bugs in, say, a reservoir sampling impl.

Cheers
Manik

On 25 May 2011, at 08:45, Galder Zamarreño wrote:

> 
> On May 24, 2011, at 1:08 PM, Dan Berindei wrote:
> 
>> On Tue, May 24, 2011 at 11:57 AM, Sanne Grinovero
>> <sanne.grinovero at gmail.com> wrote:
>>> 2011/5/24 Galder Zamarreño <galder at redhat.com>:
>>>> Guys,
>>>> 
>>>> Some interesting discussions here, keep them coming! Let me summarise what I submitted yesterday as pull req for https://issues.jboss.org/browse/ISPN-1102
>>>> 
>>>> - I don't think users can really provide such accurate predictions of the objects sizes because first java does not give you an easy way of figuring out how much your object takes up and most of the people don't have such knowledge. What I think could be more interesting is potentially having a buffer predictor that predicts sizes per type, so rather than calculate the next buffer size taking all objects into account, do that per object type. To enable to do this in the future, I'm gonna add the object to be marshalled as parameter to https://github.com/infinispan/infinispan/pull/338/files#diff-2 - This enhancement allows for your suggestions on externalizers providing estimate size to be implemented, but I'm not keen on that.
>>>> 
>>>> - For a solution to ISPN-1102, I've gone for a simpler adaptive buffer size algorithm that Netty uses for determining the receiver buffer size. The use cases are different but I liked the simplicity of the algorithm since calculating the next buffer size was an O(1) op and can grow both ways very easily. I agree that it might not be as exact as reservoir sampling+percentile, but at least it's cheaper to compute and it resolves the immediate problem of senders keeping too much memory for sent buffers before STABLE comes around.
>>>> 
>>>> - Next step would be to go and test this and compare it with Bela/Dan were seeing (+1 to another interactive debugging session), and if we are still not too happy about the memory consumption, maybe we can look into providing a different implementation for BufferSizePredictor that uses R sampling.
>>>> 
>>>> - Finally, I think once ISPN-1102 is in, we should make the BufferSizePredictor implementation configurable programmatically and via XML - I'll create a separate JIRA for this.
>>> 
>>> great wrap up, +1 on all points.
>>> BTW I definitely don't expect every user to be able to figure out the
>>> proper size, just that some of them might want (need?) to provide
>>> hints.
>>> 
>> 
>> Looks great Galder, although I could use some comments on how the
>> possible buffer sizes are chosen in your algorithm :-)
> 
> I'll ping you on IRC.
> 
>> I guess we were thinking of different things with the externalizer
>> extension. I was imagining something like an ObjectOutput
>> implementation that doesn't really write anything but instead it just
>> records the size of the object that would be written. That way the
>> size estimate would always be accurate, but of course the performance
>> wouldn't be very good for complex object graphs.
>> 
>> Still I'd like to play with something like this to see if we can
>> estimate the memory usage of the cache and base the eviction on the
>> (estimated) memory usage instead of a fixed number of entries, it
>> seems to me like that's the first question people ask when they start
>> using Infinispan.
> 
> Sure, this is something we have considered in the past, and a cache that stores everything as binary is the easiest of the use cases to provide this type of calculation.
> 
> In the case where store-as-binary is off, doing this is more complicated because even if you can marshall things at some point (i.e. at replication time), the space taken by the object in memory vs it's binary form are different.
> 
>> 
>> Cheers
>> Dan
>> 
>> 
>>> Cheers,
>>> Sanne
>>> 
>>>> 
>>>> Cheers,
>>>> 
>>>> On May 24, 2011, at 8:12 AM, Bela Ban wrote:
>>>> 
>>>>> 
>>>>> 
>>>>> On 5/23/11 11:09 PM, Dan Berindei wrote:
>>>>> 
>>>>>>> No need to expose the ExposedByteArrayOutputStream, a byte[] buffer,
>>>>>>> offset and length will do it, and we already use this today.
>>>>>>> 
>>>>>>> 
>>>>>>>> In case the value is not stored in binary form, the expected life of
>>>>>>>> the stream is very short anyway, after being pushed directly to
>>>>>>>> network buffers we don't need it anymore... couldn't we pass the
>>>>>>>> non-truncated stream directly to JGroups without this final size
>>>>>>>> adjustement ?
>>>>>>> 
>>>>>> 
>>>>>> The problem is that byte[] first has to be copied to another buffer
>>>>>> together with the rest of the ReplicableCommand before getting to
>>>>>> JGroups. AFAIK in JGroups you must have 1 buffer for each message.
>>>>> 
>>>>> 
>>>>> If you use ExposedByteArrayOutputStream, you should have access to the
>>>>> underlying buffer, so you don't need to copy it.
>>>>> 
>>>>> 
>>>>>>> You do that, yes.
>>>>>>> 
>>>>>>> However, afair, the issue is not on the *sending*, but on the
>>>>>>> *receiving* side. That's where the larger-than-needed buffer sticks
>>>>>>> around. On the sending side, as you mentioned, Infinispan passes a
>>>>>>> buffer/offset/length to JGroups and JGroups passes this right on to the
>>>>>>> network layer, which copies that data into a buffer.
>>>>>>> 
>>>>>> 
>>>>>> I don't think so... on the receiving size the buffer size is
>>>>>> controlled exclusively by JGroups, the unmarshaller doesn't create any
>>>>>> buffers. The only buffers on the receiving side are those created by
>>>>>> JGroups, and JGroups knows the message size before creating the buffer
>>>>>> so it doesn't have to worry about predicting buffer sizes.
>>>>>> 
>>>>>> On sending however I understood that JGroups keeps the buffer with the
>>>>>> offset and length in the NakReceivingWindow exactly as it got it from
>>>>>> Infinispan, without any trimming, until it receives a STABLE message
>>>>>> from all the other nodes in the cluster.
>>>>> 
>>>>> 
>>>>> Ah, ok. I think we should really do what we said before JBW, namely have
>>>>> an interactive debugging session, to clear this up.
>>>>> 
>>>>> --
>>>>> Bela Ban
>>>>> Lead JGroups / Clustering Team
>>>>> JBoss
>>>>> _______________________________________________
>>>>> infinispan-dev mailing list
>>>>> infinispan-dev at lists.jboss.org
>>>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>> 
>>>> --
>>>> Galder Zamarreño
>>>> Sr. Software Engineer
>>>> Infinispan, JBoss Cache
>>>> 
>>>> 
>>>> _______________________________________________
>>>> infinispan-dev mailing list
>>>> infinispan-dev at lists.jboss.org
>>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>>>> 
>>> 
>>> _______________________________________________
>>> infinispan-dev mailing list
>>> infinispan-dev at lists.jboss.org
>>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
>> 
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/infinispan-dev
> 
> --
> Galder Zamarreño
> Sr. Software Engineer
> Infinispan, JBoss Cache
> 
> 
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev

--
Manik Surtani
manik at jboss.org
twitter.com/maniksurtani

Lead, Infinispan
http://www.infinispan.org