Re: GWT RPC future ?

Guys,

On the heels of David's post, I was going through my application and noticed some poor RPC performance as well... much poorer than I remember. I went through my code and noticed that I'm using Longs in every object, as that's the default @Id for Objectify on AppEngine.  Is there a simple way to get the GWT-RPC to serialized these differently by default? I'd really like to keep everything as is without having to swap out everything, as the Android and AppEngine Servlets are working great.

Thanks!

E

On Tuesday, February 26, 2013 4:55:56 AM UTC-5, stuckagain wrote:

People,

Just to post some feedback on my problem, I actually found a working solution and I think there is a lesson in here that it warrant me writing back in this thread.

I managed to implement my own custom serialisation based on the flickr post. I managed to double the performance and I implemented it so that I don't use recursion to make it possible to move to an incremental approach to avoid stack overflows.

But x2 is still too long to be acceptable so I ran it through the IE9 profiler and I noticed that most of the time was actually spent in BigInteger and Long methods...

That made me remember the discussion a long time ago that longs are emulated in GWT and we should avoid them unless absolutely needed. I switched to sending a byte array for these id fields and I got a 10x speed increase (and that is with the regular GWT serialisation, not my own version). I did not realize that Longs were that costly (on IE).

So, that resolves the performance issue for me. Sure, another round of thinking is needed to avoid sending such large object trees in the first place, but this offers an acceptable solution for now - and since browsers like Chrome and FireFox (and even IE) are becoming much faster I might not even need to further improve it.

David

On Wednesday, February 13, 2013 5:31:10 PM UTC+1, stuckagain wrote:

Thomas,

I just read the article on how they improved parsing time in flickr ... really simplistic and a big surprise that the split trick is as fast as native json parsing! 

Would such an approach be usable for a generic object serialisation/deserialisation approach ? 

David

On Wednesday, February 6, 2013 5:17:33 PM UTC+1, Thomas Broyer wrote:

On Wednesday, February 6, 2013 4:37:35 PM UTC+1, stuckagain wrote:

Hi,

 

Not sure where to ask this question, but I was wondering if the GWT devs every plan to fix the inefficient GWT-RPC ?

The problem happens mostly on IE (all versions), although I assume other browsers might benefit as well since a lot of cpu cycles are wasted on things that should be trivial for a browser.

 

I had to improve multiple GWT apps that all stumble on these 3 problems:

- deserialisation is terribly inefficient - it can take many seconds to serialize small sets of data,

- on IE I can get slow script warnings

- I sometimes get stack over flows with deeply nested structures.

 

For example when I send over a tree of 10000 nodes (takes 20ms to create), it takes 5 seconds or more to deserialize. (I can give you a demo app that shows the problem)

 

I only get 2 seconds to impress my users, and I need to do quite a lot of operations besides sending the RPC.

 

I've heared the reactions multiple times: don't send soo much data over, but bytewise this is not soo much. It is highly compressible (just a few K in fact) data. We want to process complex data structures in the client, we don't want to create intermediate data structures to bypass the RPC inefficiencies.

 

There have been multiple attempts from google to write something better (DeRPC whichi is now deprecated, and RequestFactory which is very badly documented so I don't even know if I could reuse this one for generic RPC calls).

Indeed RequestFactory can be used for "generic RPC".

Have a look at http://tbroyer.posterous.com/gwt-211-requestfactory and http://tbroyer.posterous.com/gwt-211-requestfactory-part-ii

It's rather old and might be inaccurate in a few places (hasn't been updated for GWT 2.4's use of annotation-processing at compile-time, for instance).

 

 

Is it not time to start using json as the base format for GWT RPC ? I would even like to help out to get this working! It is really a pitty that somehow RPC is a selling point for GWT but in reality it often becomes the bottleneck of your application.

 

Can't we maybe put GWT RPC on the framework for request factory ?

 

One issue I also have with GWT RPC (but less pressing as the performanceissue) is the fact that it is not very friendly for mixing different client technologies. If it were a simple json REST payload (without obfuscation and lots of secret numbers) then we could easily reuse it everwhere, it would also make it soo much easier for loadtesting. Not a lot of tools support GWT RPC easily.

RequestFactory can easily be used in-process within tests, and ships with a pure-Java client (usable on Android for instance). It comes with 2 "dialects" under the same API: its own RequestFactory protocol (JSON-based) that deals with batching of method calls and sending only diffs for entities, and JSON-RPC. The server-side component only supports the former dialect though, the latter is only about using existing JSON-RPC services (such as Google APIs) from a Java or GWT app.

That said, I doubt RequestFactory would perform better for your 10000 nodes use-case (I think we can even say it will perform much worse than RPC; this can probably be improved by doing more codegen at compile-time and less reflection at runtime, but I'm not sure it'd even be better than RPC; this is mostly about the server-side though, and possibly DevMode too; it should be an all different story if you use the JSON-RPC dialect).

An alternative to RPC and RF, using (a slightly modified) JSON-RPC protocol with an RPC-like API is gwt-json-rpc, used by Gerrit: https://gerrit.googlesource.com/gwtjsonrpc/ You'll find the JAR in a Maven repo at https://gerrit-maven-repository.googlecode.com/svn/ (Gerrit itself references https://gerrit-maven.commondatastorage.googleapis.com so I think the googlecode repo is an old one; the commondatastorage.googleapis one is not browsable though so it's hard to tell which artifacts are in there). Look at the README file for details.

Finally, it's a bit old (almost 4 years old) but it should still apply as you're talking about IE: Flickr ditched JSON for a custom format for better performances, maybe you could do something similar: http://code.flickr.net/2009/03/18/building-fast-client-side-searches/

--
You received this message because you are subscribed to the Google Groups "Google Web Toolkit" group.
To unsubscribe from this group and stop receiving emails from it, send an email to google-web-toolkit+unsubscribe@googlegroups.com.
To post to this group, send email to google-web-toolkit@googlegroups.com.
Visit this group at http://groups.google.com/group/google-web-toolkit?hl=en.
For more options, visit https://groups.google.com/groups/opt_out.