On Tuesday 18 Jan 2005 19:39, Brian Stark wrote: > -----Original Message----- > From: fedora-list-bounces@xxxxxxxxxx > [mailto:fedora-list-bounces@xxxxxxxxxx]On Behalf Of Matthew Miller > Sent: Tuesday, January 18, 2005 9:52 AM > To: For users of Fedora Core releases > Subject: Re: long writes > > On Tue, Jan 18, 2005 at 09:44:09AM -0800, Brian Stark wrote: > > I'm not sure if this is the proper forum, but it seems like a good place > > to start. I am maintaining a linux cluster of servers (to date around > > 350). They have a partition configured as a RAID, the average performance > > on these servers is very good , but occasionally I will see a server go > > away for 30-40 seconds. After researching it in the code I found that the > > culprit was a "write" function call. When the call eventually completes, > > but I am writing under 100K of data so I have no idea why it would take > > that long to complete. Has anyone else seen this type of behavior? If not > > Which IO scheduler are you using? > > > -- > Matthew Miller mattdm@xxxxxxxxxx <http://www.mattdm.org/> > Boston University Linux ------> <http://linux.bu.edu/> > > Thanks for the reply, We are configured to use the deadline IO scheduling > algorithm, we made this decision after finding a similar issue on the read > side. Even with deadline our worst case on the write seems to be much > worse then the read side, an outrageous 30-40 seconds. Our servers may be > under a significant load both in memory usage and CPU, what I am mostly > interested in is an explanation. If its something we have to live with we > will work around it, but if we are doing something stupid I would rather > fix it. below is the code snippet we use to open the file: > > oflags = O_NDELAY; > oflags |= O_WRONLY; /* only writing is permitted */ > oflags |= O_CREAT; /* create the file if it doesn't exist */ > oflags |= O_TRUNC; /* truncate the file if it does exist */ > oflags |= O_NONBLOCK; /* do not block on reads and writes this will > provide us with asynchronous behavior*/ > fd = open(fn, oflags, 0666); > > The only thing that strikes me are the O_NDELAY and O_NONBLOCK flags. I > understand (now) that writes to disk will block despite these flags, but > will their usage also cause any odd behavior, or am I grasping at straws. > > Brian Do you allocate the write buffer yourself? -- Tony Dietrich ------------- War isn't a good life, but it's life. -- Kirk, "A Private Little War", stardate 4211.8
