Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!cmcl2!rutgers!sri-spam!mordor!lll-tis!ames!ucbcad!ucbvax!decvax!decwrl!spar!hunt
From: hunt@spar.SPAR.SLB.COM (Neil Hunt)
Newsgroups: comp.arch,comp.os.misc
Subject: Re: Shared libraries (Was: Re: Big Programs Hurt Performance)
Message-ID: <1420@spar.SPAR.SLB.COM>
Date: Mon, 12-Oct-87 12:16:44 EDT
Article-I.D.: spar.1420
Posted: Mon Oct 12 12:16:44 1987
Date-Received: Tue, 13-Oct-87 06:17:31 EDT
References: <6886@eddie.MIT.EDU) <2501@xanth.UUCP> <2067@sfsup.UUCP> <443@devvax.JPL.NASA.GOV> <2114@sfsup.UUCP> <439@root44.co.uk>
Reply-To: hunt@spar.UUCP (Neil Hunt)
Organization: Schlumberger Palo Alto Research - CASLAB
Lines: 53
Xref: mnetor comp.arch:2613 comp.os.misc:285

In article <439@root44.co.uk> aegl@root44.UUCP (Tony Luck) writes:
>>There is a technique [...] but this implies a
>>remarkable performance hit.
>
>On many architectures I wouldn't call this a "remarkable" performance hit
>[...] Thus a typical function call will do at least
>
>	jsr <entry point in jump table>			20 cycles
>	jmp <real address of function>			12 cycles
>	rts <back to user>				16 cycles
>
>[...] consider a real function with some
>local stack frame to set up and 3 registers to save/restore:
>
>	jsr <entry point in jump table>			20 cycles
>	jmp <real address of function>			12 cycles
>	link <set up local stack frame>			18 cycles
>	movm <save three registers>			38 cycles
>	 :
>	movm <restore three registers>			38 cycles
>	unlk <restore previous frame>			12 cycles
>	rts <back to user>				16 cycles
>
>Here the overhead is 12 cycles out of 154 i.e. just under 8% - still bad,
>bu not a complete disaster - and we still haven't counted any cycles for
>the function doing anything at all useful just a fairly typical function
>call overhead.
>
>I guess that this general overhead will not be hugely different on other
>architectures.
>
>-Tony Luck (Porting manager - Root Computers Ltd., London)

We are considering 12 cycles on a (say) 25MHz cpu, or 0.5 uS.

Now lets consider the performance when we don't have shared libraries,
and a function call hits a page fault:

	jsr <function-on-another-page>			20 cycles

	page fault interrupt				200 cycles
	disc latency delay				300000 cycles
	50000 instructions to bring in the page		750000 cycles
	1000 instructions in OS restoring state		15000 cycles

	movm <save registers>
	...


I think I will opt for the shared libraries, as soon as 
I can get my hands on SunOS releas 4.0 !

Neil/.