hprose RPC protocol speed

[danielsawan]

Hi there

I made some test in order to benchmark hprose RPC protocol and i was very surprised to see that it is slower then making an http request.

The test i made is really simple i just wrote a hello world http method that answer "hello" to an http get and wrote a client that call it 100 times.

Then i wrote a hrpose server listening on TCP having a helloworld method and call it from an hprose client 100 times.

I monitoring the execution time of both protocol and HTTP protocol win all the time expect when the number of call is above 1000.

I was exepecting that a self made protocole was quicker then HTTP.

Did i made something wrong here ?

[andot]

I don't know what's wrong, but on my computer, http get on tomcat: 2000-3000qps, and tcp hprose server: 80000-100000qps.

[danielsawan]

can you please try to put your remote method on another host then localhost ? In my test the client was on localhost but both serveur (http, hprose) was on a remote host. So the traffic had to go threw TCP like a prod context.

[andot]

If you use TCP, you can use setFullDuplex method to set the fullDuplex property to true on the client. It will be faster than the default setting.

[danielsawan]

The change is not hudge. Here is the results i get

10 requests :
duration hpro : 175899988
duration http : 82985343
HTTP WIN

100 requests :
duration hpro : 783401527
duration http : 716310436
HTTP WIN

500 requests :
duration hpro : 3416468571
duration http : 3536256223
HPROSE WIN

1000 requests :
duration hpro : 6644799171
duration http : 6926884271
HPROSE WIN

As i wrote before for a little number of request HTTP seems to be faster (real world situation) but when we loop over hundred of request then Hprose seems to be faster. I guess it is because HTTP Open & Close ressources for each call while Hprose let the connection Opened. But still in my little test HTTP seems to be faster.

[andot]

My test is 1 client, 40 threads, 25000 requests on each thread.

[wu-sheng]

@andot I think @danielsawan 's tests bring some important info. Limited requests per thread/connection seem more important to me. I suggest you should dive in to find out what is really happening in those cases.