Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.1 6/24/83; site bu-cs.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!allegra!mit-eddie!genrad!panda!talcott!harvard!bu-cs!root
From: root@bu-cs.UUCP (Barry Shein)
Newsgroups: net.emacs
Subject: Re: CCA Emacs Redisplay Bug Fix
Message-ID: <203@bu-cs.UUCP>
Date: Wed, 27-Feb-85 18:46:17 EST
Article-I.D.: bu-cs.203
Posted: Wed Feb 27 18:46:17 1985
Date-Received: Fri, 1-Mar-85 20:49:12 EST
References: <1076@godot.UUCP>
Organization: Boston Univ Comp. Sci.
Lines: 69

In a similar vein, I noticed that running CCA EMACS on our student
VAX750 (4.2bsd) seemed to be very uneven and creating a fair
amount of load. I began by studying the main loop that reads in
keystrokes and noticed the following:

	characters are read in 1 at a time with a read(0,&c,1)
	type statement

	there is polling for typeahead existence via empty(0)
	calls which invokes ioctl(0,FIONREAD,&n)

This causes a lot of calls to the system per keystroke. Read the
code and consider what occurs when all your users are using
terminals with multi-character function keys (eg. VT100s, VT200s)

Solution (maybe better put, approach): [this applies to 4.2bsd]

replace the read(0,&c,1) call with a call to a routine rdchar(0)
which buffers up as many input chars as it can (a little _filbuf/getc
routine really.) Set the character count held in the buffer
into a global. Now empty(0) becomes simply a test against this
character count which isn't as accurate but the penalty for being
wrong is much less than the penalty for so many sys calls
NOW for the thing that makes it work:

	use a select() call in rdchar to delay 1/10 of a second:
	(PSUEDO-CODE HERE)

	again:
		if(n_inbuf-- > 0) return(buf[nxtchar++])
		else {
			t->tv_sec = 0 ;
			t->tv_usec = NTENTHS * 100000 ;
			select(0,0,0,0,t) ;	/* cast these */
			if((n_inbuf = read(0,MAXBUF,buf)) <= 0) ...err
			nxtchar = 0 ;
			goto again ;
		}

	This will tend to pick up function keys in one shot,
	sometimes more than one if the person is hammering an
	arrow key or some such.

	empty becomes:

	return(!n_inbuf) ;

Just for generality's sake, I made the number of 1/10s of seconds
available as an EMACS var to the user (test for zero before the select())
Seems to help although these things are very hard to measure. That
system was definitely getting too many syscalls as judged by vmstat,
now seems better.

On a SYSV system one can just use the MINchars, MINtime features
of ~CANON for similar effect.

Yes, I still get the effect of pause while a lot of typing coming
in and then just display entire update.

I have found that 2 or 3 1/10s of a second is tolerable.

It is possible I am crocked here I suppose...anyone got a good
way to quantify the supposed advantage? I am just going by
common sense and observation which is very dangerous in
performance issues.

P.S. changes are in e_main.c and e_io.c

		-Barry Shein, Boston University