[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH 2/2] credit2: track residual from divisions done during accounting



On 27/02/13 12:48, Jan Beulich wrote:
On 27.02.13 at 13:31, George Dunlap <George.Dunlap@xxxxxxxxxxxxx> wrote:
On Wed, Feb 27, 2013 at 11:35 AM, George Dunlap <George.Dunlap@xxxxxxxxxxxxx
wrote:
On Wed, Feb 27, 2013 at 9:40 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:

On 26.02.13 at 18:08, George Dunlap <george.dunlap@xxxxxxxxxxxxx>
wrote:
@@ -271,16 +272,24 @@ struct csched_dom {

  /*
   * Time-to-credit, credit-to-time.
+ *
+ * We keep track of the "residual" time to make sure that frequent
short
+ * schedules still get accounted for in the end.
+ *
   * FIXME: Do pre-calculated division?
   */
-static s_time_t t2c(struct csched_runqueue_data *rqd, s_time_t time,
struct csched_vcpu *svc)
+static void t2c_update(struct csched_runqueue_data *rqd, s_time_t time,
+                          struct csched_vcpu *svc)
  {
-    return time * rqd->max_weight / svc->weight;
+    uint64_t val = time * rqd->max_weight + svc->residual;
+
+    svc->residual = do_div(val, svc->weight);
+    svc->credit -= val;
  }

  static s_time_t c2t(struct csched_runqueue_data *rqd, s_time_t credit,
struct csched_vcpu *svc)
  {
-    return credit * svc->weight / rqd->max_weight;
+    return (credit * svc->weight) / rqd->max_weight;
So you dropped the subtraction of svc->residual here - why?

And if indeed intended, the insertion of parentheses here could be
dropped?

Well for one I think the equation was wrong -- the residual is units of
"time", so you should have subtracted the residual after dividing by
max_weight.  (From a mathematical perspective, svc->weight /
rqd->max_weight is the ratio that changes units of "credit" into units of
"time; if we were using floating point ops I might put parentheses around
that ratio to make it more clear that's what's going on.)


Hmm, on further investigation, I think I've convinced myself that this
reasoning is wrong, and your original equation was correct.

However, the difference will only ever end up to be one nanosecond, which
(if used) will be subtracted from the credit after the reset; so I think we
might as well do without the subtraction here.
Oh, I didn't realize this would only be a nanosecond - my
understanding was that it would be up to (but not including) a
credit unit.

The "residual" will always be less than the svc->weight, which will always be less than or equal to rqd->max_weight. So adding it in to the equation before dividing by rqd->max_weight can at most increase the result by 1.

Empiraclly, I've been playing around with values in a spreadsheet to make sure the logic was behaving as one would expect, and the difference between the calculation of c2t with and without the residual was never more than 1.

 -George

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.