{[ promptMessage ]}

Bookmark it

{[ promptMessage ]}


V a possible soluon for the less typical mr jobs v

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: h may not hold. Figure 2 from [6] 27 28 Incorrect Es�mated Times v  2 Situa�ons: Second Approach v  A task’s progress rate slows down: caused the problem. v  A task’s progress rate speeds up: no problem here. v  This problem is not frequent in typical MR jobs: v  Reduce tasks are slow in their first phase then speed up. v  A possible solu�on for the less typical MR jobs: v  Account for each phase independently (per-­‐phase progress rate) 29 30 5 9/17/13 Three Approaches to Improve the Performance MR Job v  MapReduce Group Scheduler (MRG), by Kang et al. [2]. MR Job Tasks Second Approach Tasks v  A VM scheduler for scheduling the VMs on the physical machines (proposed as an extension to Xen). LATE [6]: Specula�ve Task Scheduler v  Inefficiencies of Xen: v  MR tasks are I/O-­‐bound à increase context switch P2P-­‐MR [3]: Architecture Change overhead Nodes v  Boos�ng a VM frequently can cause scheduling delay of other VMs on the same machine. Virtualized Machines v  Observa�on: For a MR job, its VMs have similar behavior. MRG [2]: VM Scheduler Physical Machines 31 MR Job Characteris�cs 32 Xen’s Credit Scheduler v  Most MR jobs are data-­‐intensive à I/O-­‐bound tasks. v  VMs are assigned credits used to schedule them on the CPU. v  Frequent I/O requests from each VM v  à Addi�onal overhead on the VMM. v  Priority order: boost > under > over > idle. v  The VM will be put in the wait queue a�er an I/O request, v  The run queue is sorted by the priori�es. and then back to the top of the run queue when the event is delivered by the driver domain. v  VMs loose some credit when de-­‐scheduled a�er allocated �me slice. v  à limits responsiveness to the other VMs. v  dom0 (special driver domain -­‐ the layer between the v  For a MR job, its VMs have similar behavior. VMs and the PM) handles all I/O processing. 33 Xen’s Credit Scheduler 34 Example v  4 domains are running on the same PM: v  If a VM request I/O, the request is passed to dom0 and v  dom0: driver domain the VM goes to the wait queue. v  dom1 and dom2 belong to a MR job, while dom3 belongs v  When dom0 is scheduled on CPU, it handles any pending to another. I/O requests. v  Upon comple�on of an I/O request, the VM is boosted in the run queue to handle the I/O. v  Scheduling dom0 too o�en incurs higher context switch overhead with fewer I/O handling per scheduling. Figure 3 from [2] v  The driver domain dom0 is scheduled a�er every de-­‐ scheduling of the VMs. 35 36 6 9/17/13 In addi�on to unfairness, they can be considered as stragglers. Example MRG Scheduler v  Idea: use the knowledge about the groups for MR jobs. v  To reduce context switch overhead: v  Instead of sor�ng the CPU run-­‐queue only using FIFO within each priority, it groups by MR job too. v  Defers dom0 to a�er a MR job group. Figure 4 from [2] If dom0 scheduling is deferred un...
View Full Document

{[ snackBarMessage ]}