View Issue Details
|ID||Project||Category||View Status||Date Submitted||Last Update|
|0001680||CentOS-4||kernel||public||2007-02-16 10:50||2007-11-05 14:10|
|Summary||0001680: Provide kernel with low interrupt timer for use in VMware|
|Description||CentOS 4 running in a VMware Server guest are observed to use inordinate amounts of CPU on the Host, even when the guest OS is completely idle.|
On my host at the moment, this varies from 15-60% depending on the speed of the host, the architecture used (x86_64 guests are worse) and the number of CPUs in the guest VM (virtual SMP exacerbates the issue). A typical average result is a constant 23% of a CPU core in system time for each CentOS 4 guest VM.
I have a 4 core Xeon Pro system with 6 guest VMs that was effectively running at 100% utilisation when there was only one guest VM doing anything but idling.
http://www.vmware.com/community/message.jspa?messageID=510526 and the related discussion threads indicate that this is an idle timer issue. Since the guest OS is not halting the virtual CPU, this will probably affect other virtualisation technologies as well (e.g. Virtual PC).
In particular it is claimed that the HZ=1000 config used in the CentOS kernel causes the issue.
In comparison, OpenSuse 10.1 guests do not appear to exhibit this high idle CPU usage, and are configured with HZ=250.
http://vmware.xaox.net/centos/4/ is a private Yum repository provided by the author of the above post, which has rebuilt CentOS kernels with HZ=100.
I've tested these on various x86_64 hosts (older Xeons and newer Xeon pros) running single and dual CPU guest VMs with x86 and x86_64 CentOS 4.4, and in all cases the CPU usage on the host has dropped significantly - below 7% per guest VM for all combinations.
We make heavy use of CentOS in our internal infrastructure (some 30+ VMs), and would appreciate if a VMware friendly kernel was available from CentOS, rather than relying on a parallel repository which could fall into disuse or get behind the main CentOS release stream.
Perhaps a kernel-vm series of kernels in centosplus could be created?
I'm not a kernel expert, so I'm not sure if there are any undesirable side-effects of dropping the interrupt timer like this.
|Tags||No tags attached.|
The packages mentioned in that url are really done badly.
We can look at providing something in a more distro friendly manner in the CentOSPlus repo, as long as hughesjr agree's with it and its something that is going to benefit a large userbase.
( btw, good bug report, I wish more people would take the time and make the effort of putting together a bug report that actually makes sense and has the relevant urls included. )
I'm not an expert in this area, but it appears that dynticks (http://kerneltrap.org/node/7749, http://kerneltrap.org/node/7568) might render this unnecessary.
If a 2.6.21+ kernel is planned for CentOS 4.5 (or 5), and it fixes the VMware issues, then that might be a better option.
|You said this could be done if it would benefit a large userbase. I just wanted to chime in and indicate that there are others who would appreciate this (myself included). Thanks.|
|Please, continue discussion in #2189.|
|2007-02-16 10:50||ultradodge||New Issue|
|2007-02-16 10:50||ultradodge||Status||new => assigned|
|2007-02-16 14:firstname.lastname@example.org||Note Added: 0004447|
|2007-02-16 14:email@example.com||Status||assigned => acknowledged|
|2007-02-26 01:08||ultradodge||Note Added: 0004477|
|2007-04-19 10:54||241comp||Note Added: 0004836|
|2007-11-05 14:09||danieldk||Relationship added||has duplicate 0002189|
|2007-11-05 14:10||danieldk||Status||acknowledged => closed|
|2007-11-05 14:10||danieldk||Note Added: 0006241|
|2007-11-05 14:10||danieldk||Resolution||open => fixed|