home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.arch
- Path: sparky!uunet!mcsun!Germany.EU.net!nixpbe!uranium!josef
- From: Josef Moellers <mollers.pad@sni.de>
- Subject: Re: Symmetric MP
- Sender: josef@nixpbe.sni.de (Moellers)
- Message-ID: <josef.711890540@uranium>
- Date: Thu, 23 Jul 1992 11:22:20 GMT
- References: <1992Jul17.183501.26138@decuac.dec.com> <147gceINN43d@early-bird.think.com> <BrKMq3.6D4@acsu.buffalo.edu> <1992Jul18.151923.211432@cs.cmu.edu>
- Organization: Siemens Nixdorf Info.Sys. AG, Paderborn, Germany
- Lines: 48
-
- In <1992Jul18.151923.211432@cs.cmu.edu> lindsay+@cs.cmu.edu (Donald Lindsay) writes:
-
-
- >pjg@acsu.buffalo.edu (Paul Graham) writes:
- >>i'm actually more interested in why it's apparently hard to do right
-
- >It is harder to do than it often looks on paper. This is partly
- >because a kernel can have a lot of locks, and a lot of possible
- >sequences of events, and the interactions get you. So, complexity,
- >and debugging.
-
- >Another major reason SMP takes time to achieve, is that almost
- >everyone seems to start from scratch. Back in the PDP-11 days, Unix
- >locking consisted of turning off interrupts whenever the kernel was
- >entered.
-
- This is not entirely true. Interrupts were (are) still enabled, only
- task-switching was (is) disabled. Therefore, critical regions could (can)
- only exist when datastructures could (can) be accessed by "normal" kernel
- functions and interrupt code. These accesses were (are) protected by
- "spl"-calls which raise(d) the interrupt level of the processor thereby
- effectively locking out interrupts FROM THIS PARTICULAR INTERRUPT SOURCE
- (and all sources that had lower priority).
- I've put past and present tenses, because a lot of single-processor UN*X
- systems still work this way nowadays.
-
- > A perfectly good decision, in its day. However, the legacy
- >is that the kernel code became a bit of a trackless wilderness, where
- >it wasn't clear what was shared and by whom. For historical reasons,
- >a large number of groups have *indepentantly* MP-ified this
- >wilderness. Well, OSF used Encore/Mach ... but who else didn't roll
- >their own? It's as bad as this mad desire to define one's own bus,
- >but with longer timelines.
-
- It's the old problem of "not invented here" combined with "If You want
- my code, You pay for it".
- In the early days of UN*X, source code flowed quite freely between
- interested parties (modtly universities and other ersearch institutes),
- because there was no commercial interest. Nowadays every single idea to
- "enhance" UN*X spells "$$$", so everyone wanting to do MP will have to
- roll their own dice.
-
- >--
- >Don D.C.Lindsay Carnegie Mellon Computer Science
- --
- | Josef Moellers | c/o Siemens Nixdorf Informationssysteme AG |
- | USA: mollers.pad@sni-usa.com | Abt. STO-XS 113 | Riemekestrasse |
- | !USA: mollers.pad@sni.de | Phone: (+49) 5251 835124 | D-4790 Paderborn |
-