Re: Plans for the SEDF scheduler

[George Dunlap <george.dunlap@eu.citrix.com>]

On 05/21/2015 01:39 PM, Dario Faggioli wrote:
> Hi everyone,
> 
> I've copied the committers, Lars as community manager, Wei as the
> release manager, George as the scheduler maintainer, RT-Xen and
> DornerWorks people, which have some interest in the field, I believe.
> Please, widen the crowd if you think it's necessary/best.
> 
> Straight to the point: the SEDF scheduler code is in a rather bad state:
> 
>  * it is buggy (tests are failing since... well, at least since when I
>    joined the community, 3 years ago!). Basically, it was not designed
>    with host and guest SMP support in mind, and has never been updated;
> 
>  * it is not used. Well, that's hard to tell, of course, but I really
>    really expect that to be the case.
> 
>  * it is useless as:
>     - it tries to be a general purpose scheduler, but we have Credit
>       (and, soon enough, hopefully, Credit2) for that
>     - for real-time we have ARINC653 (super-hard real-time), maintained
>       and supported, and RTDS (hard/soft real-time), actively
>       developed.
> 
> For these reasons, I think we should get rid of it. Actually, I've
> discovered that this was the plan in 2006 already:  :-)
> 
>  git show db51cd09d37ea44b126bb259f9392248afd768e6
>  ...
>  diff --git a/docs/src/interface.tex b/docs/src/interface.tex
>  index c9017c7..9a59840 100644
>  --- a/docs/src/interface.tex
>  +++ b/docs/src/interface.tex
>  @@ -209,8 +209,8 @@ implement timeout values when they block.
>  
>   Xen offers a uniform API for CPU schedulers.  It is possible to choose
>   from a number of schedulers at boot and it should be easy to add more.
>  -The SEDF, BVT, and Credit schedulers are part of the normal Xen
>  -distribution.  BVT and SEDF will be going away and their use should be
>  +The SEDF and Credit schedulers are part of the normal Xen
>  +distribution.  SEDF will be going away and its use should be
>   avoided once the credit scheduler has stabilized and become the default.
>   The Credit scheduler provides proportional fair shares of the
>   host's CPUs to the running domains. It does this while transparently
> 
> I'm up for sending a series to that effect, but I'd be interested in
> your opinion on what's the best way to do so.
> 
> Looking at the commit referred above, it seems like a scheduler can just
> disappear which, as far as Xen and libxc are concerned, does seem ok to
> me. Is that so?
> 
> I'm less sure of libxl, because of the API stability claims. It looks
> like this is all about (as there are not scheduler specific API
> functions, which is good) this:
> 
>  # Consistent with values defined in domctl.h
>  # Except unknown which we have made up
>  libxl_scheduler = Enumeration("scheduler", [
>      (0, "unknown"),
>      (4, "sedf"),
>      (5, "credit"),
>      (6, "credit2"),
>      (7, "arinc653"),
>      (8, "rtds"),
>      ])
> 
> The solutions I see are:
>  1. get rid of the field. This does not sound ideal from an API
>     stability point of view;
>  2. keep the field 'physically', but return warnings and errors when it 
>     is used. This is a change in behavior, so also not really stable
>     (but, hey, the scheduler is going away, the behavior will change,
>     one way or another!)

I think this is probably the best option.

>  3. RTDS has a really similar interface and, to a certain extent, a 
>     similar behavior too. Would it make sense to, upon suitable
>     warnings, of course, convert the requests within libxl, i.e.,
>     when someone asks for SEDF, we give him RTDS?
> 
> Honestly, I'd prefer 2, and I don't much like 3. However, it really
> seems that 3 gives us a way to keep most of our promises about API
> stability... Thoughts?
> 
> About domain_sched_params:
> 
>  libxl_domain_sched_params = Struct("domain_sched_params",[
>      ("sched",        libxl_scheduler),
>      ("weight",       integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_WEIGHT_DEFAULT'}),
>      ("cap",          integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_CAP_DEFAULT'}),
>      ("period",       integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_PERIOD_DEFAULT'}),
>      ("slice",        integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_SLICE_DEFAULT'}),
>      ("latency",      integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_LATENCY_DEFAULT'}),
>      ("extratime",    integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_EXTRATIME_DEFAULT'}),
>      ("budget",       integer, {'init_val': 'LIBXL_DOMAIN_SCHED_PARAM_BUDGET_DEFAULT'}),
>      ])
> 
> SEDF now uses 'weight', 'period', 'slice', 'latency' and 'extratime'. If
> considering option 3, RTDS:
>  - does not use 'weight' right now, but we can use it as an alternative
>    way to set 'budget', which is not far from the way SEDF uses it;
>  - uses 'period';
>  - does not use 'slice', but that is just a synonym for 'budget';
>  - does not use 'latency' right now; we can use it as an alternative way
>    to set 'period', but that would be a bit different from what SEDF
>    does with it;
>  - does not use 'extratime' for now, but will, at some point, in a very
>    similar way to how SEDF uses it.
> 
> Oh and, whatever route we decide to go for, is it perhaps necessary that
> we do something like deprecate it for one release (4.6), e.g., by
> printing a warning saying that <<SEDF is still here now, but is
> deprecated and you won't find it in the next release>>, and then do the
> actual removal/conversion in 4.7?

I think given that it's been "going away" for 9 years, we should
probably just remove it and put it back if people complain.  (And if
they really really can't use RTDS instead.)

 -George