improved performance of racm.nextJobsRestricted UDF. using table variable iso common table expressions in single table. somehow MSSQL started performing very bad, even if only ~200 queries were pending.

Author: glemson

components/java/racm/src/main/resources/db/migration/V34__Alter_nextJobsRestricted_UDF.sql

@@ -0,0 +1,71 @@
+CREATE or ALTER function [racm].[nextJobsRestricted](@compmUUID varchar(64), @timeout real, 
+        @interval real, @maxNum integer,@maxPerUser integer) 
+returns @rt table(
+	  jobId bigint,submitterId bigint,
+	  computeDomainId bigint, submitTime datetime,
+	  usageWeight float,numQueued integer, numStarted integer, ranking integer) 
+as
+begin
+
+declare @computeDomainId bigint, @currentDate datetime = getDate()
+
+select @computeDomainId = computeDomainId
+  from COMPM
+ where uuid=@compmUUID
+;
+
+declare @prev table (submitterId bigint, numJobs integer, numQueued integer
+   ,  numStarted integer, totTime float, usageWeight float)
+
+insert into @prev
+select dj.submitterId
+,      count(*) as numJobs
+,      sum(case when dj.status between 2 and 4 then 1 else 0 end) as numQueued
+,      sum(case when dj.status = 8 then 1 else 0 end) as numStarted
+,      sum(case when dj.finishedTime is null then @timeout -- should really be cast(datediff(second,startedTime, getDate()) as float)
+                else cast(datediff(second,dj.startedTime, dj.finishedtime) as float)
+				end ) as totTime
+,      sum((case when dj.finishedTime is null then @timeout -- should really be cast(datediff(second,startedTime, getDate()) as float)
+                else cast(datediff_big(millisecond,dj.startedTime, dj.finishedtime) as float)
+				end )/datediff_big(millisecond,dj.startedTime,@currentDate)) as usageWeight
+  from compm c , job dj 
+ where c.uuid = @compmUUID
+   and dj.computeDomainId=c.computeDomainId
+   and dj.status > 1 -- PENDING
+   and dj.startedTime >= dateadd(second,-@interval,getDate())
+ group by submitterid
+
+declare @pending table (id bigint, submitterId bigint, computeDomainId bigint
+       , submitTime datetime,rankSubmitted integer)
+
+insert into @pending
+select dj.id, dj.submitterId, dj.computeDomainId, submitTime
+,      rank() over (partition by submitterId order by submitTime) as rankSubmitted
+  from compm c , job dj 
+ where c.uuid = @compmUUID
+   and dj.computeDomainId=c.computeDomainId
+   and dj.status=1 -- PENDING
+
+;
+with final as (
+select j.Id as jobId,j.submitterId, j.computeDomainId, j.submitTime  -- without TOP cannot do an order by in a function
+,      isnull(p.usageWeight,0) as usageWeight
+,      isnull(p.numQueued,0) as numQueued
+,      isnull(p.numStarted,0) as numStarted
+,      rank() over (order by rankSubmitted,usageWeight,submitTime) as ranking
+  from @pending j left outer join @prev p
+  on p.submitterId = j.submitterid
+  where j.rankSubmitted <=@maxNum
+    and j.rankSubmitted+isnull(p.numQueued,0)+isnull(p.numStarted,0)<=@maxPerUser
+)
+insert into @rt
+select * -- jobId,submitterId, computeDomainId, submitTime , usageWeight, numQueued, numStarted, ranking
+  from final
+ where ranking <= @maxNum 
+--   and rankSubmitted+numQueued+numStarted<=@maxPerUser -- to counter possibility COMPM gets single user too many times in internal queue
+
+  return 
+end
+
+
+