Subject: Re: oprofile archive patch updated for oprofile cvs - msg#00020

Previous Message by Date: click to view message preview

accounting for I/O wait time

Hello, I was wondering if it is possible to account for the time the CPU spends waiting for I/O using oprofile, separate from the CPU idle time. I will really appreciate if someone could give any pointers in this regard. Thanks, Piyush ------------------------------------------------------- This SF.Net email is sponsored by: IBM Linux Tutorials Free Linux tutorial presented by Daniel Robbins, President and CEO of GenToo technologies. Learn everything from fundamentals to system administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click

Next Message by Date: click to view message preview

Re: accounting for I/O wait time

Piyush Shivam wrote: Hello, Hi, At Duke University? You are just down the road from me. I was wondering if it is possible to account for the time the CPU spends waiting for I/O using oprofile, separate from the CPU idle time. OProfile is pretty much oriented toward CPU-bound applications. OProfile is great for finding which areas of code are CPU hogs. You can sometimes find I/O bottlenecks with OProfile from the routines that the kernel spends time in, e.g. processor spends a lot of time in driver code to copy data between device and memory. However, if the CPU isn't doing anything, then OProfile doesn't record much. The sampling OProfile uses doesn't show where the processor sits idle. This is quite possible with I/O when the processor is waiting for the device to complete some action, e.g. a disk drive head seek or waiting for an interrupt signalling that a network packet has arrived. What kind of I/O are you talking about? Mass storage or networking? Or some custom device? Do you have some simple benchmarks setup for repeatable experiment and to keep the complexity managable? I will really appreciate if someone could give any pointers in this regard. If you are looking at I/O for specific devices, something that measures interverals of time required to get from one place in the code to another might be more useful. Something like Linux Trace Toolkit (LTT) that provided probes to specific places in the kernel might give more information, http://www.opersys.com/LTT/. I haven't use LTT, so this is just a guess. -Will ------------------------------------------------------- This SF.Net email is sponsored by: IBM Linux Tutorials Free Linux tutorial presented by Daniel Robbins, President and CEO of GenToo technologies. Learn everything from fundamentals to system administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click

Previous Message by Thread: click to view message preview

Re: accounting for I/O wait time

Thanks for your reply. Yes, I am at Duke. We are trying to do some bottleneck analysis for some interesting applications which have bursty I/O behavior. We are doing a whole system profile where a client connects to the nfs server over Ethernet. The goal is to understand the behavior of all the resources of the system so that one can see whats going on behind the scenes. The applications do have a lot of I/O happening. Thanks for the pointer you gave me. I will have a look. -Piyush On Mon, 12 Apr 2004, William Cohen wrote: > Piyush Shivam wrote: > > Hello, > > Hi, > > At Duke University? You are just down the road from me. > > > I was wondering if it is possible to account for the time the CPU spends > > waiting for I/O using oprofile, separate from the CPU idle time. > > OProfile is pretty much oriented toward CPU-bound applications. OProfile > is great for finding which areas of code are CPU hogs. > > You can sometimes find I/O bottlenecks with OProfile from the routines > that the kernel spends time in, e.g. processor spends a lot of time in > driver code to copy data between device and memory. However, if the CPU > isn't doing anything, then OProfile doesn't record much. The sampling > OProfile uses doesn't show where the processor sits idle. This is quite > possible with I/O when the processor is waiting for the device to > complete some action, e.g. a disk drive head seek or waiting for an > interrupt signalling that a network packet has arrived. > > What kind of I/O are you talking about? Mass storage or networking? Or > some custom device? Do you have some simple benchmarks setup for > repeatable experiment and to keep the complexity managable? > > > I will really appreciate if someone could give any pointers in this > > regard. > > If you are looking at I/O for specific devices, something that measures > interverals of time required to get from one place in the code to > another might be more useful. Something like Linux Trace Toolkit (LTT) > that provided probes to specific places in the kernel might give more > information, http://www.opersys.com/LTT/. I haven't use LTT, so this is > just a guess. > > > -Will > ------------------------------------------------------- This SF.Net email is sponsored by: IBM Linux Tutorials Free Linux tutorial presented by Daniel Robbins, President and CEO of GenToo technologies. Learn everything from fundamentals to system administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click

Next Message by Thread: click to view message preview

Re: oprofile archive patch updated for oprofile cvs

John Levon wrote: On Mon, Apr 12, 2004 at 05:35:47PM -0400, Will Cohen wrote: I have updated the old oprofile archive patch to work with the current version of oprofile in cvs. It adds an "archive:" specifier and a new Nice and simple... does that project on savannah (I always forget the name) have a generic copy file routine we can take? "cp" is surprisingly painful to get right, never mind fast. It wasn't too painful given the current organization of oprofile. I will see if savannah has suitable code for a generic copy. That is where most of the pain (and FIXMEs) is right now. Other than that just the usual comments: cstyle, moving generic code into generic libs etc. Looks good to me. I will see what I can do to clean this up this week to make it suitable for merging into the OProfile cvs. Any thoughts on when the next release of OProfile will be? -Will ------------------------------------------------------- This SF.Net email is sponsored by: IBM Linux Tutorials Free Linux tutorial presented by Daniel Robbins, President and CEO of GenToo technologies. Learn everything from fundamentals to system administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click