CS655-Alsaraan-MapReduce

V if it is less than a threshold more likely to

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: l a�er dom2, the I/O requests can be processed by a single scheduling. à Reduce context switch overhead. 37 Sor�ng Example 38 MRG Scheduler v  Possible problem of deferring dom0: v  A group of VMs are not reques�ng any I/O. v  Causes delay to the other I/O pending VMs. v  Solu�on: v  Predict the I/O request likelihood of a VM. v  It is based on running �me of its last scheduling. v  If it is less than a threshold -­‐> more likely to request I/O next scheduling. Figure 5 from [2] 39 MRG Scheduler 40 Hierarchical Scheduling v  To improve fairness: v  First level: v  The CPU �me allocated for a VM depends on its credit v  Allocate credits for groups: and its MR group credit. v  The weight of a group is the sum of each of its VMs credits. v  It provides fair CPU �me to each group, and then to each v  User allocated credits. VM belonging to a group à hierarchical scheduling. v  VM credits are normalized as v  The scheduler keeps track of the running �me of each group Cg × (Ci / ∑ C j ) v  Second Level: v  Within each group’s allo�ed �me, only VMs of this group can run. v  Allocate CPU to to VMs in a group using their credits. v  à tasks in the same group will progress around the same rate. 41 42 7 9/17/13 Example For Group 1: CPU alloca�on for the group: 2/6 of overall CPU �me. Within group sharing ra�o: 2:1. CPU alloca�on for VMs: 2/9 and 1/9 of overall CPU �me. (2/3).(2/6)=(4/18)=(2/9) Hierarchical Scheduling v  Possible problem: v  It is possible that all the VMs in a group are blocked in the wai�ng list a�er reques�ng too many I/O. v  à Wasted CPU �me. v  Solu�on: v  A �meout counter is associated with each group. v  Start the �meout when all VMs within the group are blocked. v  If the �meout expires and they are s�ll all blocked à the group yields the CPU to the next group. Figure 6 from [2] 43 Se�ng the Timeout Counter 44 Se�ng the Timeout Counter v  Credit-­‐remaining ra�o (CRR) for each group: v  Se�ng it too low: v  à The scheduler switches between groups quickly. CRR = v  à Pending I/O are not given a chance to complete. ∑VM _ remaining _ credit Total _ group _ credit v  The frac�on of CPU credits remaining for the group. v  Affects group-­‐level fairness. v  If CRR is close to 1 à the group is far from using all its credits. While if it is close to zero à the group has used most of its credits. v  Se�ng it too large à the system is underu�lized. v  tw = wakeup latency between the �me when all VMs were v  The scheduler must select it to balance these trade-­‐offs. blocked and the �me when any of them was boosted. (ini�alized with 1 ms). v  The �meout: t = tw . CRR 45 46 Hierarchical Scheduling v  Addi�onal extensions: Third Approach v  The current running group is selected as the group of the VM in the head of the running queue. v  The scheduler must only select VMs from the gr...
View Full Document

{[ snackBarMessage ]}

Ask a homework question - tutors are online