1. ECF: an MPTCP Scheduler to Manage Heterogeneous Paths
Yeon-sup Lim1, Erich M. Nahum1, Don Towsley2, and Richard Gibbens3
1 IBM T.J. Watson Research Center
2 University of Massachusetts Amherst
3 University of Cambridge
ACM CoNEXT, Incheon, Korea, Dec. 2017

2. Introduction
Multi-Path TCP: simultaneously utilizes multiple interfaces
Packets can be scheduled to each subflow according to several policies
How do these scheduling policies affect MPTCP performance?
App
subflow-1
Wi-Fi
Socket
MPTCP
subflow
subflow … …
WiFi
LTE
subflow-2
Cellular

3. MPTCP Default Scheduler
Default scheduler sends packets over available subflow with lowest RTT
Connection level send buffer
Connection level
subflow level send buffer
MPTCP
Scheduler
subflow 1 with RTT1
MPTCP scheduler
subflow
subflow … .
subflow 2 with RTT2

4. What is ideal average bit rate given bandwidths?
MPTCP default scheduler provides ideal aggregate bandwidth? - Video Streaming Case
Dynamic Adaptive Streaming over HTTP (DASH)
Video : 22min, 6 resolutions, 5s chunks
Required Bandwidth (bit rate) for each resolution
Resolution
144p
240p
360p
480p
760p
1080p
Bit Rate (Mbps)
0.26
0.64
1.00
1.60
4.14
8.47
What is ideal average bit rate given bandwidths?
e.g. 0.3 Mbps WiFi Mbps LTE > 8.47 Mbps (1080p)
All chunks served with 1080p Mbps*
* Depends on ABR schemes, but ABR basic goal is to achieve this average bit rate

5. MPTCP default scheduler provides ideal aggregate bandwidth
MPTCP default scheduler provides ideal aggregate bandwidth? - Video Streaming Case
Measured average bit rate ratio to ideal
0.3Mbps WiFi & 8.6Mbps LTE
Ideal = > 8.47 (1080p) → 8.47
Measured = 2.30
Ratio = 2.30 / 8.47 = 0.28
0.3Mbps WiFi & 0.3 Mbps LTE
Ideal = ≈ 0.64 (240p) → 0.6
Measured = 0.55
Ratio = 0.55 / 0.6 = 0.92
Close to 1 (darker) -> Average bit rate is close to ideal

6. Motivation Default scheduler causes idle period of fast subflow
< 𝑥 𝐶𝑊𝑁 𝐷 𝑓 times tx
Connection level send buffer
<RTTf
RTTf 𝑥
idle … …
CWND full
Pkt for next GET
Subflow with smaller RTT
RTTs …
CWND available
Subflow with larger RTT
send

7. Motivation Waiting for fast subflow can complete transfer earlier
𝑥 𝐶𝑊𝑁 𝐷 𝑓 times tx
Connection level send buffer
<RTTf
RTTf 𝑥 … …
CWND full
Pkt for next GET
Subflow with smaller RTT
RTTs …
CWND available
Subflow with larger RTT
send

8. Motivation Why do these fast subflow idle periods matter?
Available bandwidth of fast subflow is not utilized during idle periods
CWND frequently restarts by RFC 5681
Another bandwidth loss in fast subflow until CWND becomes sufficient value
0.3Mbps WiFi & 8.6Mbps LTE

9. Earliest Completion First (ECF) Scheduler
If fastest subflow is available, just use it
Otherwise, check subflow selected by default scheduler (second fastest subflow):
If 𝑅𝑇 𝑇 𝑓 + 𝑥 𝐶𝑊𝑁 𝐷 𝑓 ×𝑅𝑇 𝑇 𝑓 <𝑅𝑇 𝑇 𝑠
i.e., waiting for fastest subflow can complete tx. earlier than using slower one at this moment
Do not use second fastest subflow
If not, Use second fastest subflow
i.e., sender has enough large # of packets in send buffer to efficiently utilize both subflows

10. Experimental Setup Implemented in MPTCP Linux Kernel (version 0.89)
Mobile device (Google Nexus 5) downloads several types of traffic using WiFi and LTE
Streaming, HTTP file download, and Web page download
Compares to Default, DAPS1, and BLEST2 schedulers
Scenarios
Controlled in-lab experiments with various conditions
Evaluation in the wild
[1] N. Kuhn et al., “DAPS:Intelligent delay-aware packet scheduling for multipath transport”, IEEE ICC’14
[2] S. Ferlin et al., “BLEST: Blocking estimation-based MPTCP scheduler for heterogeneous networks”, IFIP Networking’16

11. In-lab experiments – Streaming w/ fixed bandwidths
Measured average bit rate relative to ideal (5 runs)
Default Scheduler
ECF
Close to 1 (darker) -> Average bit rate is close to ideal

12. In-lab experiments – Streaming w/ fixed bandwidths
ECF
BLEST
DAPS
Improves for some cases, but worse for many others
Obtains some gains, but ECF is still better

13. In-lab experiments – Streaming w/ fixed bandwidths
LTE CWND Trace (0.3Mbps WiFi and 8.6Mbps LTE)
Default’s LTE CWND frequently back to initial value
BLEST works better than Default, but
More often back to initial than ECF

14. In-lab experiments – Web browsing
Replicated CNN home page (as of 9/11/2014) – 107 Web objects in page
HTTP persistent connection is enabled (6 connections)
5Mbps WiFi & 5Mbps LTE
(Similar paths)
1Mbps WiFi & 10Mbps LTE
(Heterogeneous paths)

15. Evaluation in the Wild - Streaming
DASH server at Washington DC
Streaming Client at Cafe using public town WiFi in Amherst, MA
9 runs over two days
* Sorted by avg. WiFi RTT
small RTT difference
- Default and ECF will yield similar performance

16. Evaluation in the Wild - Streaming
similar Avg RTTs (70ms)
WiFi has extremely large RTT (≈1sec).

17. Evaluation in the Wild – Web browsing
Replicated CNN home page (as of 9/11/2014) at WDC server
Object Download Completion Times from 30 runs
99.9% of object downloads complete in
17 seconds with ECF
30 seconds with Default

18. Summary
Investigated reason of MPTCP performance degradation in presence of path heterogeneity
Proposed and implemented new subflow scheduler (ECF) considering completion time
Details for implementation in
Evaluated ECF scheduler with several types of traffic
ECF is never worse than default scheduler and substantially better with path heterogeneity
More experimental results in paper
HTTP Download workload, Out-of-order delay analysis, …

19. Thank you!
Questions?

20. In-lab experiments – Streaming w/ fixed bandwidths
Backup
In-lab experiments – Streaming w/ fixed bandwidths
ECF is working with more than two subflows?
Experiments using two subflows over WiFi and two over LTE
Regulated subflows over each interface to evenly provide designated bandwidth
For 8.6 Mbps LTE, each subflow over LTE provides 4.3Mbps bandwidth
0.3 Mbps WiFi and [0.3, 8.6] Mbps LTE

21. In-lab experiments – Streaming w/ random bandwidths
Backup
In-lab experiments – Streaming w/ random bandwidths
Average throughput during streaming
Gains depend on how often heterogeneity happens
Note: x axis is chunk index, not time

22. In-lab experiments – HTTP downloads
Backup
In-lab experiments – HTTP downloads
File downloads using wget for single file of which size is 65KB-1MB
In case of small file downloads (128KB)
Download time is very short. Second flow is rarely used
No difference between default and ECF
Ratio of ECF download time to Default
(128KB Downloads)
More Red: ECF worse
More Blue: ECF better
White: no difference

23. In-lab experiments – HTTP downloads
Backup
In-lab experiments – HTTP downloads
In case of large file downloads (256KB to 1MB)
ECF works better with path heterogeneity
Note:
Only one idle period is expected at the end of transfer
Significantly large file → Long transfer time → Small effect of idle period → Small gain from ECF
256KB
512KB
1MB