|
|
| Re: IO scheduling [message #6668 is a reply to message #6610] |
Thu, 21 September 2006 06:43   |
Vasily Tarasov
Messages: 1345
Registered: January 2006
|
Senior Member |
|
|
Hi,
I'll try to explain you what is in Vserver concerning IO scheduling and why it isn't now in OpenVZ.
There is cfq scheduler in Linux kernel, that allows to assign IO priority to the process. It supports three classes:
real time (rt)
best effort (be)
idle class (i)
Within rt class and be class are 8 levels of priority. The more level is - more time for input/output particular process has. Additional information can be found for example at http://www.mjmwired.net/kernel/Documentation/block/ioprio.tx t
So what do they do in Vserver?
When user sets certain IO priority to the context, Vserver framework just sets this IO priority to all processes in context! And that's all that they do, but this isn't right. Just look at these example:
1st context: 3 processes - priority be:4
2nd context: 1 process - priority be:6
So user expects that 2nd context has more IO bandwidth, but this isn't true, 'cause 1st context has more processes! And the more processes 1st context has more IO bandwidth it has.
Some time ago there were patches to do the same in OpenVZ, but do we need such implementation? 
To create more sophisticated and true IO scheduling more investigation is necessary. Also there is also a big problem, 'cause pages can be written to the block device, when information about process isn't available any more...
HTH,
vass.
|
|
|
|
|
|
|
|
| Re: IO scheduling [message #6838 is a reply to message #6679] |
Mon, 25 September 2006 07:05 |
HaroldB
Messages: 61
Registered: June 2006
|
Member |
|
|
Hello. The ability for a VE to utilize all of the disk i/o bandwidth in a system seems to be a very big problem. Has anyone investigated a project called CKRM?
"If you want a way to assign io priorities without relying on process inheritance and (re)nice you might find CKRM, with it's cfq-based IO controller, useful.
| Quote: |
Basically you create a set of classes that group tasks and give an
appropriate share of IO performance to tasks in that class. As processes get created CKRM will assign tasks to the IO classes based on a set of rules."
|
ref:
http://ckrm.sourceforge.net/
http://www.gatago.com/linux/kernel/14683383.html
Seems like if each CKRM "class" was a openvz VE, this could be a nice framework for limiting and more importantly guaranteeing disk i/o bandwidth per VE. Quoted from the CKRM patch:
| Quote: |
Resource allocations for a class is controlled by the parameters:
guarantee: specifies how much of a resource is guranteed to a class. A
special value DONT_CARE(-2) mean that there is no specific
guarantee of a resource is specified, this class may not get
any resource if the system is runing short of resources
limit: specifies the maximum amount of resource that is allowed to be
allocated by a class. A special value DONT_CARE(-2) mean that
there is no specific limit is specified, this class can get all
the resources available.
total_guarantee: total guarantee that is allowed among the children of this
class. In other words, the sum of "guarantee"s of all children
of this class cannot exit this number.
max_limit: Maximum "limit" allowed for any of this class's children. In
other words, "limit" of any children of this class cannot exceed
this value.
|
[Updated on: Mon, 25 September 2006 07:09] Report message to a moderator |
|
|
|
| Re: IO scheduling [message #6908 is a reply to message #6668] |
Wed, 27 September 2006 11:04 |
wfischer
Messages: 38
Registered: November 2005
Location: Austria/Germany
|
Member |
|
|
Another question regarding io scheduler:
To me it seems that OpenVZ uses anticipatory io scheduler as default io scheduler, according to the info in /var/log/messages after booting OpenVZ kernel (on a CentOS 4.4 host):
Sep 22 09:14:04 wc1 kernel: Using anticipatory io scheduler
I'm not an expert on io scheduling, but I heard that anticipatory io scheduler is mainly useful for desktop machines and for servers deadline or cfq schedulers should be used.
Could you give a short explanation why anticipatory io scheduler is used as default io scheduler in OpenVZ?
Thanks,
best wishes,
Werner
added remark: I just noticed that the anticipatory io scheduler has different default values in OpenVZ than in vanilla kernel (read_expire 10 instead of 125, read_batch_expire 10 instead of 500) - according to the suggestion on http://bugzilla.kernel.org/show_bug.cgi?id=5900#c1
Werner Fischer, Developer of a Virtuozzo-out-of-the-box-cluster solution at Thomas-Krenn.AG
[Updated on: Wed, 27 September 2006 11:14] Report message to a moderator |
|
|
|
|
|
|
|
| Re: IO scheduling [message #6913 is a reply to message #6912] |
Wed, 27 September 2006 12:17 |
wfischer
Messages: 38
Registered: November 2005
Location: Austria/Germany
|
Member |
|
|
Thanks a lot for the info and the fast reply, it helps me a lot.
@HaroldB: I'm using OpenVZ Kernel 2.6.8-022stab078.14, so that is the reason why anticipatory is default according to the explanation from Vass.
best regards,
Werner
Werner Fischer, Developer of a Virtuozzo-out-of-the-box-cluster solution at Thomas-Krenn.AG
|
|
|
|
OpenVZ Forum
Members
Search
Help
Register
Login
Home
Well, yes, there is no possibility to directly limit IO now, not in OpenVZ not in Vserver 
