From owner-freebsd-questions Wed Aug 2 15:03:59 1995 Return-Path: questions-owner Received: (from majordom@localhost) by freefall.cdrom.com (8.6.11/8.6.6) id PAA02600 for questions-outgoing; Wed, 2 Aug 1995 15:03:59 -0700 Received: (from dyson@localhost) by freefall.cdrom.com (8.6.11/8.6.6) id PAA02591 ; Wed, 2 Aug 1995 15:03:58 -0700 From: John Dyson Message-Id: <199508022203.PAA02591@freefall.cdrom.com> Subject: Re: 2.0.5 Eager to go into swap To: serges@umr.edu Date: Wed, 2 Aug 1995 15:03:57 -0700 (PDT) Cc: freebsd-questions@freefall.cdrom.com In-Reply-To: from "serges@umr.edu" at Aug 2, 95 03:40:35 pm X-Mailer: ELM [version 2.4 PL24] Content-Type: text Content-Length: 2949 Sender: questions-owner@FreeBSD.org Precedence: bulk > > Well I dont know what those "mechanisms" are and I (unlike the original > poster) cant compare 2.0.5 with a previous version of FreeBSD, but I can > say that I experience excessive swapping on my system. I have a 486DX50 > with 20 megs of core and ~20 megs of swap space. With a minimal X > desktop with Netscape and 2 xterms running I can *easily* exhaust the virtual > memory on my system! This is ofcourse, after running Netscape (or xv) for > a long time (> 1hour continuous use). I usually have to kill the server > and restart things. > That is the *exact* problem that I have been talking about!!! It appears that you don't have enough swap space also. Add up the virtual space for all of your processes (esp the processes that there is only one running, like X) and you might find that you have more than 20MBs... You CAN run with less swap than RAM, but the system currently doesn't have code to make that as efficient as it could be. (Especially with long-lived processes like the X server, which then permanently takes the swap space until it exits (or deallocates RAM.)) At the point of filled swap space, the system will tend to page against .text too much. That might be modified in a future release. Also, as swap space gets filled, it becomes fragmented -- and that can slow things down also. > I can understand that the server grows, thanks to memory leaks in X11 > itself, but I dont understand why the 2.0.5 system begins swapping > *before* the core resources are even 50% in use. Coming from a SVR4 How do you know? When the Xserver has lots of sparse memory allocated, it can take a huge working set -- maybe :-). I have noticed that once I run one of those *evil* applications, my Xserver appears to almost permanently want more resident memory. That tells me that the resident set that it demands is much larger. Also, the swap space that it keeps is larger. (The OS really doesn't know that the memory is no longer needed unless the app frees it.) > background on Intel machines, I find it a hassle to watch as all of my > xterms slowly redraw themselves as they are swapped back in to core - > after I briefly switch back from another application. > I hate that too. > And I dont think its solely the fault of the clients and their real or > virtual sizes. I feels more like a scheduling design at fault here. But > Im speaking as a user (of Unices on many different platforms), and not a > kernel programmer. > It is 95% a working set problem, that no OS can easily solve. I am a user (and have been a kernel maintainer of SVR4) and can say that given the same client mixes and VM demands -- the FreeBSD VM is much faster. (I have benchmarked and compared carefully.) BTW, I don't want to give XFree a bumb rap -- it is really neat. It is just that it appears that part of the problem is that the FreeBSD malloc might be clashing with XFree??? John dyson@root.com