From owner-freebsd-current Tue Nov 30 9:28:22 1999 Delivered-To: freebsd-current@freebsd.org Received: from apollo.backplane.com (apollo.backplane.com [216.240.41.2]) by hub.freebsd.org (Postfix) with ESMTP id CE3EE14E73 for ; Tue, 30 Nov 1999 09:28:03 -0800 (PST) (envelope-from dillon@apollo.backplane.com) Received: (from dillon@localhost) by apollo.backplane.com (8.9.3/8.9.1) id JAA25821; Tue, 30 Nov 1999 09:28:02 -0800 (PST) (envelope-from dillon) Date: Tue, 30 Nov 1999 09:28:02 -0800 (PST) From: Matthew Dillon Message-Id: <199911301728.JAA25821@apollo.backplane.com> To: Thomas Stromberg Cc: freebsd-current@FreeBSD.ORG Subject: Re: -CURRENT crash under high exec() loads.. (vm_map_insert?) References: <38430B2D.2E7A5F0D@rtci.com> <199911300115.RAA13409@apollo.backplane.com> <3843DDF3.793414E7@rtci.com> Sender: owner-freebsd-current@FreeBSD.ORG Precedence: bulk X-Loop: FreeBSD.ORG :> that doesn't work, and there's any chance of getting a kernel dump, :> try getting a kernel dump. Make sure you use a debug (compiled -g) :> kernel. :> : :While I can't say I truly see the correlation, you definitely know more :about how it works then me. I've cvsupp'd and upped it to 1000. One :question I have, is what exactly -are- pv's? PV's are a mistake, but nobody has figured out a better way to do it. PV's are essentially FreeBSD's way of managing page tables. They associated a physical page with an entry in a process's page table and allow the VM system to track which physical pages are mapped (and where), and which are not. This in turn reduces the overhead of many of the algorithms in the VM system and allows the system to scale. However, it costs memory to do this and it has an achilles heel - shared address spaces can exhaust the pool without exhausting physical memory. :The machine itself isn't under too much load during the testing, and :definitely does not used a lot of shared memory. During one of the tests :yesterday (c89) it was doing ~8000 execs per minute, with a load of :0.23. I've also got plenty of RAM left over. Here is what it currently :looks like: Yah. I don't know if that's it or not, but whenever systems crash that deep into the VM system the first thing we try is to up the number of PV's. :Free :Swap: 786M Total, 22M Used, 764M Free, 2% Inuse : :This is basically X, Netscape, and 5800 exec()'s a minute. : :I've got debugging in my kernel, and I'll send you full results if and :when it crashes again. :thomas r. stromberg smtp://tstromberg@rtci.com Both X and netscape have large run sizes. Any fork/exec from either of those programs will create a temporary spike in resource use. I could be totally wrong but it's definitely suspect and worth trying. -Matt Matthew Dillon To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message