Re: CGIDEV2 performance versus Net.Data

Responses below.

Mel Rothman
CGIDEV2 Author
IBM eServer iSeries Custom Technology Center (iCTC)
http://www-1.ibm.com/servers/eserver/iseries/service/ctc/
Rochester, Minnesota


"Nathan M. Andelin" wrote:
>
> > From: "Mel Rothman" <mel@rothmanweb.com>
>
> > As Leif points out, what's really being used is virtual memory.
>
> I don't understand this.  Or how it relates to the issue of scalability.
>
The point is that the amount of virtual storage allocated by activated programs
doesn't directly affect performance much.  What does is when the demand on real
storage significantly exceeds the available real storage.

A couple of simplified examples to illustrate the point.  It's really a lot more
complicated because paging is a lot more granular than I'm using in the example
and is influenced by things like activation levels, run priorities, time slice
intervals, etc.  The system manages storage much better than one might infer
from these examples.

First, 10 activated programs, each running in a separate job, each having
allocated 1MB of virtual storage = 10 MB of virtual storage.  At a given
instant, if 6 are active, only 6MB of real storage are required.  If the system
has 8 MB of real storage available for programs, 2 of the 4 non-running programs
will have been paged out but the 6 running programs should run fine with no
additional paging required.  The other 2 non-running programs are ready to run
without paging because their pages are in real storage.  If 3 more suddenly
start running (total 9), 9 MB of real storage are required, 1MB more than
available.  So, 1 of the 9 running programs would have to be paged out (perhaps
one that was already paged out) and wait for real storage to become available at
which time it would run.

Using the same 8Mb of available memory, if there are 100 activated programs each
requiring 1MB of memory, but most of the time no more than 8 are running, paging
will occur as one stops and another starts (assuming the one just starting was
paged out), but there won't be thrashing.  If 20 are trying to run most of the
time, there will be a lot more paging, probably thrashing, and performance will
suffer.  In a case like this with the HTTP server, it would be important to
"tune" the number of running programs to some number perhaps a little above 8,
but significantly below 20.

So, the amount of required virtual storage is not directly related to the amount
of required real storage.  In the second example, if the number of activated
programs was 500 or 1000, but still no more than 8 ran at a time, performance
would be similar to what it would be with 100 activated programs or 50 activated
programs.  So, the critical factor is the real storage required by running
programs.  In reality, the more real storage, the better.

> > this is a problem when activity gets so high that excessive
> > paging occurs.
>
> If CGI programs are active, but not running, would that increase the
> occurance of paging?

Yes and no, as illustrated in the second simple example above when 8 or fewer of
the 100 programs are running concurrently.  Yes, paging occurs when a program
that had been paged out starts running, but the system can handle this kind of
normal paging easily.  The more the real storage required by running programs
exceeds available real storage, the more paging increases and the harder it
becomes for the system to keep up.
>
> > As previously pointed out, this can and should be controlled
> > by throttling the server.
>
> That makes sense.  It's an imposed bottleneck.

Yes, but it's much better than allowing thrashing.  This is commonly called
performance tuning or optimizing.

> > It's not clear whether you approach supports a remote
> > application server.
>
> Not presently.  But it wouldn't take much to add remote routing via data
> queue.

That's what I thought.  Note that sockets are much faster than data queues.  At
some point, you might want to offer your customers the choice of data queues or
sockets.  The iCTC's sockets server is excellent.  Read all about it at:
http://www-1.ibm.com/servers/eserver/iseries/service/ctc/details/a2_ssrv.htm

> The Relational-Web version of my sample application uses about 80% less CPU
> time than the Easy400 version, and about 90% less CPU time than the Net.Data
> version.

This is a nit but you previously said Easy400 and Net.Data used 80% and 90% more
than Relational-Web.  If so, then Relational-Web would use 44% and 47% less than
Easy400 and Net.Data.

> I'm not sure how significant those percents are.  Is it important to people?
> Perhaps not.  A tool can only do so much.  It's really up to the programmer
> to create efficient software with the tools provided.

Agreed.

>
> Thanks,
>
> Nathan M. Andelin
> www.relational-data.com