From: Santos, J. R. G <jos...@hp...> - 2007-07-19 19:02:37
|
Will, Thanks for pointing this out. I will work on a patch to fix this By the way, it would be nice to have the xenoprof patches merged into Oprofile CVS tree. John Levon is not happy with the way passive domain samples are stored now and I don't have the bandwidth (and knowdledge of user level Oprofile code) to do the right fixes. This has prevented the patches to be merged. It would be nice if we can find a volunteer to do the right fixes and have the code merged. I would be happy to assist anyone who volunteers for the task. What we need is a way of separating samples from different domains and change to the tools to understand this new information. Thanks Renato=20 >> -----Original Message----- >> From: xen...@li...=20 >> [mailto:xen...@li...] On Behalf Of=20 >> William Cohen >> Sent: Tuesday, July 17, 2007 11:47 AM >> To: xen...@li... >> Subject: [Xen-devel] xenoprof support patch issue for oprofile 0.9.3 >>=20 >> Oprofile 0.9.3 released yesterday. The patches for xenoprof=20 >> are not currently in the upstream oprofile. I noticed that=20 >> one of the defines in the xenoprof patches for oprofile has=20 >> a define DOMAIN_SWITCH_CODE in daemon/opd_interface.h which=20 >> conflicts with SPU_PROFILING_CODE. This conflict will need=20 >> to be worked out. >>=20 >>=20 >>=20 >> -Will >>=20 >> _______________________________________________ >> Xen-devel mailing list >> Xen...@li... >> http://lists.xensource.com/xen-devel >>=20 |
From: William C. <wc...@re...> - 2007-07-20 14:44:31
|
Santos, Jose Renato G wrote: > Will, > > Thanks for pointing this out. > I will work on a patch to fix this > By the way, it would be nice to have the xenoprof patches merged into > Oprofile CVS tree. > John Levon is not happy with the way passive domain samples are stored > now and I don't have the bandwidth (and knowdledge of user level > Oprofile code) to do the right fixes. This has prevented the patches to > be merged. It would be nice if we can find a volunteer to do the right > fixes and have the code merged. I would be happy to assist anyone who > volunteers for the task. What we need is a way of separating samples > from different domains and change to the tools to understand this new > information. > > Thanks > > Renato Hi Renato, Definitely want to avoid having separate patches to avoid just this type of problem. Getting patches upstream is the way to go on that. Do you have a pointer to the email discussion about the passive domain sample storage? Passive means that the domain is not taking the samples and storing them in the domain, right? One of the things that I have noticed on my previous experiments with the xenoprof is there are quite a few steps to setup xenoprof. The user has to do setup on each domain. Ideally it would be nice to be able to just do one setup in the host domain and then have that information be propagated to the guest domains. Similarly the analysis would be nice to get one over all picture of all the domains. Clusters have a similar issue; one doesn't want to manage machine individually. However clusters may have different hardware for the various nodes in the cluster. With xen the only difference could be the software. -Will |
From: Santos, J. R. G <jos...@hp...> - 2007-07-20 16:45:21
|
=20 >> -----Original Message----- >> From: xen...@li...=20 >> [mailto:xen...@li...] On Behalf Of=20 >> William Cohen >> Sent: Friday, July 20, 2007 7:44 AM >> To: Santos, Jose Renato G >> Cc: xen...@li...;=20 >> opr...@li... >> Subject: Re: [Xen-devel] xenoprof support patch issue for=20 >> oprofile 0.9.3 >>=20 >> Santos, Jose Renato G wrote: >> > Will, >> >=20 >> > Thanks for pointing this out. >> > I will work on a patch to fix this >> > By the way, it would be nice to have the xenoprof patches=20 >> merged into=20 >> > Oprofile CVS tree. >> > John Levon is not happy with the way passive domain=20 >> samples are stored=20 >> > now and I don't have the bandwidth (and knowdledge of user level=20 >> > Oprofile code) to do the right fixes. This has prevented=20 >> the patches=20 >> > to be merged. It would be nice if we can find a volunteer=20 >> to do the=20 >> > right fixes and have the code merged. I would be happy to assist=20 >> > anyone who volunteers for the task. What we need is a way of=20 >> > separating samples from different domains and change to=20 >> the tools to=20 >> > understand this new information. >> >=20 >> > Thanks >> >=20 >> > Renato >>=20 >> Hi Renato, >>=20 >> Definitely want to avoid having separate patches to avoid=20 >> just this type of problem. Getting patches upstream is the=20 >> way to go on that. >>=20 >> Do you have a pointer to the email discussion about the=20 >> passive domain sample storage? I will forward the email discussion I had with John on a private message >> Passive means that the domain=20 >> is not taking the samples and storing them in the domain, right? >>=20 Right! This makes profiling guests in kernel and xen mode much=20 simpler than with active domains and avoids the complexities you mention below. With passive domains there is no need to run oprofile on guests as dom0 process all guest samples. The disadvantage is that you cannot assign=20 user level samples to the right binaries in the guest. >> One of the things that I have noticed on my previous=20 >> experiments with the xenoprof is there are quite a few steps=20 >> to setup xenoprof. The user has to do setup on each domain.=20 >> Ideally it would be nice to be able to just do one setup in=20 >> the host domain and then have that information be propagated=20 >> to the guest domains. Similarly the analysis would be nice=20 >> to get one over all picture of all the domains. Clusters=20 >> have a similar issue; one doesn't want to manage machine=20 >> individually. However clusters may have different hardware=20 >> for the various nodes in the cluster. With xen the only=20 >> difference could be the software. >>=20 Yes, profiling multiple guests (as active domains) is somewhat complex. But, it should not be difficult to have some wrapper scripts to coordinate oprofile execution on multiple domains, using possibly ssh, hiding the complexities from the user. =20 Renato =20 >> -Will >>=20 >> _______________________________________________ >> Xen-devel mailing list >> Xen...@li... >> http://lists.xensource.com/xen-devel >>=20 |
From: William C. <wc...@re...> - 2007-07-26 14:29:26
|
Hi Renato, I have munged the xenoprof patch to apply cleanly to the user-space oprofile. The patch don't address the larger changes with handling passive domains. The new oprofile-0.9.3-2.fc8 rpm built with this patch can be found at: http://koji.fedoraproject.org/koji/buildinfo?buildID=11990 The xen kernel space will need to change the defines to correctly match. DOMAIN_SWITCH_CODE and LAST_CODE have changed to: +#define DOMAIN_SWITCH_CODE 13 +#define LAST_CODE 14 -Will |
From: Santos, J. R. G <jos...@hp...> - 2007-07-30 23:11:56
|
Will Xenoprof patches for oprofile 0.9.3 (to enable passive domain profiling) are now available at http://xenoprof.sourceforge.net. A required patch for the kernel is also available at the same location. I think the kernel patch should NOT be included into the public Xen tree at this time, as this will break compatibility with oprofile 0.9.2 that users are currently using. I think it is better to have users wanting to upgrade to oprofile 0.9.3 to apply an extra patch to the kernel rather than breaking compatibility for all current users. I think we should wait until Xen dom0 support gets into mainline linux to do that (or at least until xenoprof patches are included in an official oprofile release). Regards Renato =20 >> -----Original Message----- >> From: William Cohen [mailto:wc...@re...]=20 >> Sent: Thursday, July 26, 2007 7:29 AM >> To: Santos, Jose Renato G >> Cc: xen...@li...;=20 >> opr...@li... >> Subject: Re: [Xen-devel] xenoprof support patch issue for=20 >> oprofile 0.9.3 >>=20 >> Hi Renato, >>=20 >> I have munged the xenoprof patch to apply cleanly to the=20 >> user-space oprofile.=20 >> The patch don't address the larger changes with handling=20 >> passive domains. The new oprofile-0.9.3-2.fc8 rpm built with=20 >> this patch can be found at: >>=20 >> http://koji.fedoraproject.org/koji/buildinfo?buildID=3D11990 >>=20 >> The xen kernel space will need to change the defines to=20 >> correctly match.=20 >> DOMAIN_SWITCH_CODE and LAST_CODE have changed to: >>=20 >>=20 >> +#define DOMAIN_SWITCH_CODE 13 >> +#define LAST_CODE 14 >>=20 >>=20 >> -Will >>=20 |