Re: SOAP performance data and open benchmarks

Hi Ken,

>Research at Rutgers measured the latency of Java RMI, CORBA, and various
>implementations of SOAP (Microsoft, Apache SOAP, SOAP::Lite, etc.). RMI and
>CORBA were in the ballpark with TCP setup time. For a mix of tests, the SOAP
>implementations were slower by factor of 12 to 239:
>http://www.caip.rutgers.edu/TASSL/Papers/p2p-p2pws02-soap.pdf

I am familiar with their benchmark test which were presented at the CCGrid
conference last summer. I talked to the authors at the conference and they
acknowledged that the toolkits were still under development and one should
not take the tests as a definite fact. Actually, most toolkits had a large
start-up latency (e.g. >12ms). Also, most toolkits did not support HTTP 1.1
keep-alive back then. Keep-alive reduces latencies significantly.

>This paper discussed RMI and SOAP benchmarks using linked links and serialized
>double arrays:
>http://www.sc2000.org/techpapr/papers/pap.pap261.pdf

I am very familiar with this work. The authors have a follow-up paper of
this and I gave you the URL earlier. The paper concludes that over 90% of
the total time is spend serializing and deserialing doubles with strtod and
sprintf calls! This makes my point: you'll never get the performanced of
Java RMI for certain messages, despite company PR claims. Our gSOAP paper
reviews alternative representations for large arrays of doubles, thereby
omitting the need for the XML double float representations.

Comparing Java RMI (and CORBA) to SOAP amounts to comparing apples with
pears. SOAP/XML's power is platform and language neutrality. The price is
performance, although this is really an issue of communication bandwidth in
most cases.

- Robert

You can read messages from the SOAP archive, unsubscribe from SOAP, or subscribe to other
DevelopMentor lists at http://discuss.develop.com.