1 Better Scalable Algorithms for Broadcast Scheduling Ravishankar Krishnaswamy Carnegie Mellon University Joint work with Nikhil Bansal and Viswanath Nagarajan (IBM T. J. Watson Research Lab)
Feb 16, 2016
1
Better Scalable Algorithms for Broadcast Scheduling
Ravishankar KrishnaswamyCarnegie Mellon University
Joint work with Nikhil Bansal and Viswanath Nagarajan (IBM T. J. Watson Research Lab)
Outline
• Motivation, Problem Definition
• Existing Results, Our Results
• A Weaker Approximation/ Analysis
• Conclusion
2
Motivation: Client-Server System
Clients Server
Page A at time 1
Page B at time 1
Page A at time 2Page C at time 3
Page A at time 3
Page APage BPage CPage Abroadcast
Motivation: Formalizing
• Consider a server which has n unit-sized pages– Requests for these pages arrive online, over time– At each time slot, we can broadcast one page• All pending requests for that page are satisfied
• How do we schedule to minimize average response time of requests
4
Online Broadcast Scheduling
• Input– A collection of n pages– A request sequence arrives online
• Request r: arrival time a(r), requested page p(r)
• Output– A broadcast of pages, one at a time
• Objective Function– Minimize Average Response Time– Minimize Maximum Response Time– …
This Talk
A Concrete Example
Instance has 3 pages
ABC
A B C
A B
B
A BC
Total Response Time: 1 + 2 + 3 + 3 + 3 = 12
A
Total Response Time: 2 + 3 + 1 + 1 + 1 = 8
Existing Results (Average Response Time)
• In the offline setting• O(log2n)-approximation algorithm [BCS06]
• In the online setting• very strong lower bounds if no speed-up• (2+є) speed-up, O(1/є2)-competitive[EP09] • (1+є) speed-up, O(1/є11)-competitive [IM10]
(works only for unit-size pages)
Our Results
8
A very simple (1+є) speed, O(1/є3)-competitive online algorithm. Can be extended to the setting when the pages have non-uniform sizes
High Level Idea
9
Consider “Fractional” Relaxation of Broadcast Scheduling
Get (1+є) speed, O(1/є2) competitive online algorithm
Design an online rounding algorithm, with further O(1/є) loss in obj. function
Fractional Relaxation
• At each time slot, we can broadcast multiple pages, each to extent xpt
– Such that
• A request r is satisfied at the first time b(r) when
• Minimize
10
High Level Idea
11
Consider “Fractional” Relaxation of Broadcast Scheduling
Get (1+є) speed, O(1/є2) competitive online algorithm
Design an online rounding algorithm, with further O(1/є) loss in obj. function
Algorithm (with weaker guarantee)
• Round Robin– Known to give online algorithms with good
competitive ratio for other scheduling problems assuming factor of 2 speed-up
– What about broadcast scheduling?– Naïve algorithm is bad• Does not differentiate pages with many outstanding
requests and those with 1 request
12
Algorithm (with weaker guarantee)
• Round Robin: Possible Fix– Round robin over requests!
At any time, schedule each outstanding request to the same extent.
• Illustration
13
ABC
A B
A: 1/3B: 1/3C: 1/3
A: 2/4B: 1/4C: 1/4
A: 2/5B: 2/5C: 1/5
A: 1/4B: 2/4C: 1/4
A: 1/3B: 1/3C: 1/3
…
Algorithm (with weaker guarantee)
• Round Robin: Possible Fix– Round robin over requests!
At any time, schedule each outstanding request to the same extent.
• Can we show anything for this algorithm?– Edmonds and Pruhs showed it is 4-speed O(1) competitive
• We show that fractionally, it is 2-speed O(1) competitive– Later round it to get integer schedule.
14
Analysis
15
• Resort to an amortized analysis• Define a potential function Φ(t) which is 0 at t=0 and t=• Show the following:
– At any request arrival,
ΔΦ ≤ 0 – At all other times,
15
Will give us a β-competitive online algorithm
For our Problem
• Define
• rank(r) is sorted order of requests w.r.t arrival times (most recent has highest rank)
• z(r,t) is the amount of time the online algorithm will dedicate towards request r, in the future, i.e. after time t
16
Analysis Continued
• New request arrival– It belongs to NA(t) and NO(t)– Does not appear in potential function– No change in value
17
Analysis Continued
• Running Condition: Consider [t-1, t)• Opt schedules a page and finishes some requests• These terms will now appear in the potential function.• How much increase will it cause?
– The sum of the z(r,t) over all these requests is at most 1– Total increase is at most NA(t)
18
We’re golden if NO(t) is even a tiny fraction of NA(t)
Analysis Continued
• Assume most unfinished requests are completed by OPT• Hope that Φ(t) goes down enough.
19
(i) We make progress on all jobs(ii)Each job’s z value goes down by 1/NA(t)(iii)Total decrease is NA(t)/2 * 1/NA(t) * 2
(iv)Left hand side is non-positive!Speed-Up
High Level Idea
20
Consider “Fractional” Relaxation of Broadcast Scheduling
Get (1+є) speed, O(1/є2) competitive online algorithm
Design an online rounding algorithm, with further O(1/є) loss in obj. function
Rounding: One Slide Overview
• Consider the fractional algorithm’s output• Let request r be fractionally completed at time b(r)• Enqueue element <r, b(r) – a(r)>• At any time, choose request with least width and
display corresponding page. Wipe out all outstanding requests for page p(r)
21
Suppose a request was forced to wait for too much time.Then many other requests for different pages all having smaller width.
Too much mass packed fractionally. A contradiction.
Thank You
Summary + Open Question
• Near-optimal algorithm for broadcast scheduling– Consider “fractional relaxation”– Give good algorithm for fractional problem– Give rounding scheme for integral problem
• But algorithm depends on є– Not fully-scalable – Can we get one such algorithm which works for all є?
22