Path: utzoo!utgpu!jarvis.csri.toronto.edu!clyde.concordia.ca!uunet!mcsun!hp4nl!sara5!toon
From: toon@vax1.sara.nl
Newsgroups: news.software.anu-news
Subject: Re: PCA stats on a NEWS-reading session
Summary: How to speed up ANU-news
Keywords: pca, speedup, anunews
Message-ID: <48@vax1.sara.nl>
Date: 30 Jan 90 09:19:56 GMT
References: <8910310355.AA11260@uunet.uu.net> <1794.25546a51@esat.kuleuven.ac.be> <17198@kuhub.cc.ukans.edu> <18387.25c07443@merrimack.edu>
Reply-To: toon@sara.nl
Organization: SARA, Academic Computing Services Amsterdam (NL)
Lines: 45

In article <18387.25c07443@merrimack.edu>, rand@merrimack.edu writes:
> 
> I just got PCA and decided to give it a go. Since most use of NEWS, and
> hence where performance is an issue, is centered around reading, I profiled
> a typical (for me) reading session. When reading I tend to:
> 
> a) Read registered groups sequentially
> b) When reading a group I scan down the directory with a keypad key
>    defined as DOWN 18 and read items of interest (20-30%)
> c) After viewing all desired items I mark them all as read with a
>    keypad key and do a SELECT/NEW bound to a keypad key
> d) goto a)

.. follows a lot of numbers ...

> (Did anyone actually read all of this? ;-)

A collegue and I did a subset of this already in November and subsequently
mailed the results to Geoff (we were using ANU-news 5.8 then). We never
heared of it since. In our experiments ga_exact_name was the culprit, so
we changed it to do the strcmp inline, which sped up this routine by about
50 %; by now the problem is in the index evaluation ga[i]->grp_name,
which we can't change without redesigning ANU-news fundamentally.

Of course, in our case also the RMS calls were important. I even mentioned
to Geoff changing the NEWS.GROUPS file into a global section and mapping
it once, at the beginning of a news session. The solution to use RMS
global buffers is of course a better one, because it doesn't require
rewriting ANU-news.

BTW, before all this, I'd already modified the ANU-news sources to use
&(constant expression) wherever Geoff used c$rfi(constant expression) in
function argument lists. See for this (non-ANSI) trick the VAX C manual
(page 14-11 for VAXC 2.4). Therefore in our measurements c$alloc_tmp
doesn't show up. With this modification you save up to four function
calls per occurrence (very frequent in screen management routine calls)
and a lot of (expensive) heap space management (ANSI folks, are you
listening ?).
-- 

Toon Moene, SARA - Amsterdam (NL)
Internet: TOON@SARA.NL

Disclaimer: Don't need one. Real Programmers (as opposed to Quiche Eaters)
            are not afraid to claim.