1. Finishing Flows Quickly with Preemptive Scheduling
Speaker: Yu, Ye
The slides are based on the paper by Chi-Yao Hong, et. al, SIGCOMM 2012
Some figures origins from their presentation.

2. outline Motivation PDQ solution to flow scheduling Evaluation
Discussion

3. Datacenter Networks

4. Latency 1ms latency = 1% reduce of sales.
100ms latency = 0.2% number searches
2.2 seconds faster / page
= 60 M more download / year

5. LOW LATENCY LOW Latency Datacenters? Finish Flows Earlier
The LAST of flows == final result
Meet Flow Deadline
User-facing applications, Latency Goal

6. partition aggregate model
associated component deadlines in the parentheses.

7. Today’s transport protocols
TCP / RCP/ ICTCP / DCTCP:
Fair Sharing
Divide link bandwidth equally.
Fail to reduce flow completion time.

8. WHAT is TCP TCP slow start TCP fast recovery additive increase
multiplicative decrease
Host A
one segment
RTT
Host B
time
two segments
four segments

9. WHAT IS RCP Rate Control Protocol
RCP is an adaptive algorithm to emulate Processor Share : a router divides outgoing link bandwidth equally
Rate is picked by the routers based on queue size and aggregate traffic
Router assigns a single rate to all flows
Requires no per-flow state or per-packet calculation

10. Fairness damages completion time
Flow Fa,Fb,Fc arrives at the same time, with size = 1,2,3 and deadline = 1,4,6
Fair share,
FC time = (3+5+6)/3 = 4.67
D3 for order BAC
FC time = (2+4+6)/3 = 4
Shortest Job First/ Earliest Deadline First
FC time = (1+3+6)/3 = 3.33

11. Satisfy request by Order
D3 depends on flow order
D3 satisfies as many flows as possible in the order of their arrival,
Request rate = flow size / time until deadline.
Satisfy request by Order

12. The PDQ Solution Preemptive Distributed Quick
Schedule by flow criticality.
Criticality: relative priority of flows.
Scheduling discipline.
Preemptive : relating to the purchase of goods or shares by one person or party before the opportunity is offered to others.

13. PDQ’s scheduling diciplines
EDF: earliest deadline first
Optimal for flow deadlines.
SJF: shortest job first
Optimal for mean flow finish time.
EDF+SJF:
Give preference to deadline flows.
Policy based:
Manually allocate priority of flow.

14. challenges. Decentralizing scheduling discipline
More mice than elephant.
Switching between flows seamlessly
Hard to full utilize bandwidth
Prioritizing flows using FIFO tail-drop Queues
FIFO Queue length limited

15. outline Motivation PDQ solution to flow scheduling Evaluation
Discussion

16. PDQ protocol - overview

18. PDQ protocol-PDQ sender-1
SYN / TERM packet for initialization and termination.
Resend after timeout.
sender maintains info for in-flight packets:
Current Sending Rate (Rs)
ID of switch who paused it (Ps)
Deadline (Ds)
Expected flow transmission time (Ts)
Inter-probing time (Is)
Measured RTT (RTTs)

19. PDQ protols-sender-2 Sender sends package with rate Rs
If Rs = 0, Send a probe packet heartbeatly.(scheduling header without data)
When ACK arrives, update Rs (ACKinfo: accept/pause)

20. Pdq protols-sender -early-termination
Sender TERMNINATES a flow when it cannot meet its deadline. Whenever:
Deadline is past.
Remaining flow transmission + time > deadline
Flow is paused , and time + RTT> deadline

21. PDQ protols-switch Let the most critical flow complete asap.
Critical flows preempt others to achieve the highest possible sending rate
1) maintain state about each flow
2) Compute Rate Feedback
a) flow controller to decide witch flows to send
b) rate controller to determine Rate

22. pdq protocol-switch-state
Maintains flow states on each link
<Rate, P, Deadline, expected Time,RTT>
Pi: flow i is paused by switch Pi
Store 2K of them, most critical ones. K is number of Current Sending Flow.

23. pdq protocol-FLOW Control
Whenever a Switch receives ACK/data, ACCEPT or PAUSE a flow
Pause: inform others flow f is Paused.
Switch who receives ACK-Pause i removes i from its own states
Accept: calculate available bandwidth
Other Switch who receives ACK-accept i updates state i

24. algorithm recv data/ack flow f
1) if f is paused by other Switch, remove it from my list.
2) if f is not in my list:
Try to add f into my list, if can not ,pause f
3) if (w = min(aviliableBW ,Rf) > 0 ):
Accept f
Otherwise pause f

25. Flow-control-3 optimization
Dampening
If switch accepted a flow, then in a short period of time he can not accept other new flows.
Early starting
Suppressed probing

26. Early start seamless schedule

27. Suppressed probing Sender may send probe packages too often.
Flow info If : tell the sender of f that you should send probe every If*RTT.
If is maintained by switches , by calculation of average finish time of all flows and rank of f

28. pdq protocol-rate control
Control the total sending rate of its accepted flows.
Maintains variable C to compute range of Rate.
reserves BW for early started flows
C = Full_BW- Queue_size/(K*RTT)

29. outline Motivation PDQ solution to flow scheduling Evaluation
Discussion

30. Evaluation setting: TRAFFIC
Deadline-constrained flows:
Time sensitive : ~20ms
Short message : 2KB~200KB
Goal: Application Throughput = percentage of flows meets their deadlines
Deadline-unconstrained flows:
100~1000KB
Goal: average flow completion time

31. Evaluation SETTING: topology

32. query aggregation:
All senders initiate at the same time to the same receiver.
Optimal: one scheduler control all transmission with no delay.
maximize application throughput:
sort by EDF, and then uses a dynamic programming

33. The Deadline-unconstrained case

34. Seamless flow switching
Five flow (~1MB) comes at the same time

35. an elephant flow and 50 short flows starting from 10ms

36. Impact of network scale

37. outline Motivation PDQ solution to flow scheduling Evaluation
Discussion

38. Fairness?

39. Other concerns Does it require rewriting APP? Deployment?
PDQ paused appears like TCP slow,
The transport connection stays open.
Deployment?
Hosts: between IP and transport layer
Switch: modify hardware/software, O(k)

40. Thank you!
Q&A