X-Original-To: pgsql-performance-postgresql.org@localhost.postgresql.org Received: from localhost (unknown [200.46.204.2]) by svr1.postgresql.org (Postfix) with ESMTP id EF4D6D1E154 for ; Wed, 17 Mar 2004 05:17:48 +0000 (GMT) Received: from svr1.postgresql.org ([200.46.204.71]) by localhost (neptune.hub.org [200.46.204.2]) (amavisd-new, port 10024) with ESMTP id 67678-04 for ; Wed, 17 Mar 2004 01:17:46 -0400 (AST) Received: from sss.pgh.pa.us (unknown [192.204.191.242]) by svr1.postgresql.org (Postfix) with ESMTP id 1E8FDD1DF84 for ; Wed, 17 Mar 2004 01:17:45 -0400 (AST) Received: from sss2.sss.pgh.pa.us (tgl@localhost [127.0.0.1]) by sss.pgh.pa.us (8.12.11/8.12.11) with ESMTP id i2H5HeYr021012; Wed, 17 Mar 2004 00:17:41 -0500 (EST) To: Joe Conway Cc: "Matthew T. O'Connor" , pgsql-performance , Josh Berkus Subject: Re: rapid degradation after postmaster restart In-reply-to: <4057D8BD.4080902@joeconway.com> References: <4052661D.5060505@joeconway.com> <14884.1079150555@sss.pgh.pa.us> <4055F343.6070600@joeconway.com> <405614CA.5030101@zeut.net> <405689A4.4080406@joeconway.com> <16735.1079414723@sss.pgh.pa.us> <4057D8BD.4080902@joeconway.com> Comments: In-reply-to Joe Conway message dated "Tue, 16 Mar 2004 20:49:01 -0800" Date: Wed, 17 Mar 2004 00:17:40 -0500 Message-ID: <21011.1079500660@sss.pgh.pa.us> From: Tom Lane X-Virus-Scanned: by amavisd-new at postgresql.org X-Archive-Number: 200403/232 X-Sequence-Number: 6088 Joe Conway writes: > I have tested Tom's original patch now. The good news -- it works great > in terms of reducing the load imposed by vacuum -- almost to the level > of being unnoticeable. The bad news -- in a simulation test which loads > an hour's worth of data, even with delay set to 1 ms, vacuum of the > large table exceeds two hours (vs 12-14 minutes with delay = 0). Since > that hourly load is expected 7 x 24, this obviously isn't going to work. Turns the dial down a bit too far then ... > The problem with Jan's more complex version of the patch (at least the > one I found - perhaps not the right one) is it includes a bunch of other > experimental stuff that I'd not want to mess with at the moment. Would > changing the input units (for the original patch) from milli-secs to > micro-secs be a bad idea? Unlikely to be helpful; on most kernels the minimum sleep delay is 1 or 10 msec, so asking for a few microsec is the same as asking for some millisec. I think what you need is a knob of the form "sleep N msec after each M pages of I/O". I'm almost certain that Jan posted such a patch somewhere between my original and the version you refer to above. regards, tom lane