RE: [xml-dev] When parsing speed matters (was Re: [xml-dev] No XML Binar

[Richard Salz <rsalz@us.ibm.com>]

This has been an interesting discussion.

There's recently been a "merge" of accelerators to cover both appliances 
(e.g., DataPower) and accelerator cards (e.g., Tarari).  They're 
different.

A card *might* be useful for parsing/tokenization, for example. I'm 
skeptical -- the latency cost of pushing the data across the PCI bus is 
more likely to hurt overall system performance, unless you're very careful 
and tune (or ideally build) your software to leverage it.  You can't just 
drop an accelerator into Apache or IIS with a JAR or DLL and expect things 
to just go faster. Notably, Tarari's benchmarks don't talk about parsing, 
but instead mention things like schema validation and XPath evaluation -- 
where all the data goes *into* the card, but very little comes back, and 
where your main application might still have to do all the heavyweight XML 
work.

(For those who are curious, the case is very different for crypto cards 
accelerating RSA operations. In those cases, the data being pushed is 
small and fixed-size (usually 1K or 4K) and the compute time greatly 
exceeds the latency.)

An appliance, however, is generally used to do "higher level" XML work 
such as implementing the WS-* stack, validating incoming credentials 
against the organization's security system, or even things like converting 
SOAP/HTTP to Cobol Copybook/MQ. It can be a pretty compelling argument 
when you tell someone that they can do full XML security, schema 
validation, threat prevention, etc., with absolutely no additional cost to 
their application deployment.  (In fact, if it was terminating all the SSL 
connections, there will be a speedup since the transient clients are 
replaced by a persistant HTTPS/1.1 connection.)  And as a hardened 
appliance, you put it in your DMZ.

Hope this helps.

        /r$

--
STSM
Senior Security Architect
DataPower SOA Appliances