From mboxrd@z Thu Jan 1 00:00:00 1970 From: John Kacur Subject: Re: Calculate average latencies on the fly Date: Mon, 21 Nov 2016 19:32:41 +0100 (CET) Message-ID: References: Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Cc: Clark Williams , Linux RT Users To: Piotr Gregor Return-path: Received: from mail-wm0-f45.google.com ([74.125.82.45]:35390 "EHLO mail-wm0-f45.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754009AbcKUScw (ORCPT ); Mon, 21 Nov 2016 13:32:52 -0500 Received: by mail-wm0-f45.google.com with SMTP id a197so162962307wmd.0 for ; Mon, 21 Nov 2016 10:32:51 -0800 (PST) In-Reply-To: Sender: linux-rt-users-owner@vger.kernel.org List-ID: On Sat, 19 Nov 2016, Piotr Gregor wrote: > Hi John, Clark, > > Recently I observed jitter of > 100 000 during tests of virtualized system. > In that case 30 years would change into 0.3 year which is not as much. > The jitter depends on the hardware/software you test, what if someone > deliberately wishes to measure jitter of more than this or he is just > out of luck and gets huge values? > Also, the limit will change if the type of average changes, but if it > is calculated in the way I proposed, it will always work. > Apart from this, since the benign possibility exists - why not to fix this > if the fix is easy. > I already explained why. Everytime you do a floating point long division you are potentially accumulating small errors. Something is terribly wrong if you are getting a diff of greater than 100,000, if that really is true you need to fix it right away instead of run cyclictest for a long period of time. John>