V also not true in pay as you go sengs 19 assumpons

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: from 1/3 to 1 quickly à increases the average progress score. v  Ex: network, and disk I/O. v  Needless specula�on can reduce throughput. v  Also not true in “pay-­‐as-­‐you-­‐go” se�ngs. 19 Assump�ons Invalidated 20 Assump�ons Invalidated v  Example: 30% of reducers finish, while the others are v  The assump�on that a straggler’s likelihood depend on s�ll in the copy phase. its progress score. v  Average progress score = (0.3).(1) + (0.7).(1/3) ~= 53% v  New and fast tasks can be chosen over old and slow tasks. v  The threshold jumped from (33 – 20 = 13%) to (53 – 20 = v  It should also consider progress rate. 33%) à all the remaining 70% of the reducers are considered for specula�on. à task slots will be filled from an arbitrary selec�on. v  Average progress score limit is 100% à Threshold = 80%. à Tasks with progress score > 80% can never be considered for specula�on. v  True stragglers may have not been chosen. v  Network overloaded with unnecessary copying. 21 LATE Scheduler 22 LATE Scheduler v  Goal: reduce the MR job response �me. v  Determining which nodes are slow: v  A node’s total performed work = sum of progress scores v  Approach: for both completed and in-­‐progress tasks. 1.  Priori�ze the specula�ve tasks based on their es�mated v  If total < SlowNodeThreshold à it is a straggler. finish �me. v  To reduce resources cost: 2.  Schedule the tasks on fast nodes (avoid stragglers). v  specula�veCap: a cap on the number of running v  Es�ma�ng a task’s finish �me: specula�ve tasks at a �me. v  Progress rate = progressScore / T v  Use the SlowTaskThreshold as an addi�onal condi�on for running specula�ve tasks. v  Remaining �me = (1 – progressScore) / progressRate v  If there are only fast tasks that are running à no need to v  Assump�on: the task’s progress rate is constant (may not speculate them. hold). 23 24 4 9/17/13 LATE Scheduler LATE Scheduler v  Algorithm: v  Se�ng parameters: (by prac�ce) v  When a node asks for a task v  Specula�veCap = 10% of available task slots. v  slowNodeThreshold = 25th percen�le of node progress. v  If the number of running specula�ve tasks < specula�veCap v  If the node’s total progress > slowNodeThreshold v  slowTaskThreshold = 25th percen�le of task progress rates. v  Rank running tasks (that are not currently speculated, and have run for more than 1 minute) by es�mated remaining �me v  Choose the highest-­‐ranked task with progress rate < slowTaskThreshold, then send it the asking node. v  Note: it does not consider data locality such as in Hadoop. 25 LATE Scheduler 26 Incorrect Es�mated Times -­‐ Example v  Some incorrect es�mated �mes: v  The es�mated remaining �me for a task was based on the progress rate: (1 – ProgressScore) / ProgressRate. v  This assumed constant progress rate per task, whic...
View Full Document

Ask a homework question - tutors are online