SLIDE 1
Detecting if LTE is the Bottleneck with BurstTracker
Arjun Balasingam, Manu Bansal, Rakesh Misra, Kanthi Nagaraj, Rahul Tandra, Sachin Katti, Aaron Schulman
Diagnosing Poor Streaming Quality
2
360p
Detecting if LTE is the Bottleneck with BurstTracker Arjun - - PowerPoint PPT Presentation
Detecting if LTE is the Bottleneck with BurstTracker Arjun - - PowerPoint PPT Presentation
Detecting if LTE is the Bottleneck with BurstTracker Arjun Balasingam , Manu Bansal, Rakesh Misra, Kanthi Nagaraj, Rahul Tandra, Sachin Katti, Aaron Schulman Diagnosing Poor Streaming Quality 360p 2 Diagnosing Poor Streaming Quality 360p 2
SLIDE 2
SLIDE 3
Diagnosing Poor Streaming Quality
2
360p
SLIDE 4
Diagnosing Poor Streaming Quality
2
🤕
360p
My video quality is only 360p (1.5 Mbps). The cellular downlink must be slow.
SLIDE 5
Diagnosing Poor Streaming Quality
2
🤕
360p
My video quality is only 360p (1.5 Mbps). The cellular downlink must be slow.
SLIDE 6
Diagnosing Poor Streaming Quality
2
🤕
But a speed test says the downlink is 10 Mbps. What’s going on?!
360p
My video quality is only 360p (1.5 Mbps). The cellular downlink must be slow.
😡
SLIDE 7
Diagnosing Poor Streaming Quality
2
🤕
But a speed test says the downlink is 10 Mbps. What’s going on?!
We can not conclusively determine if the cellular downlink is the bottleneck.
360p
My video quality is only 360p (1.5 Mbps). The cellular downlink must be slow.
😡
SLIDE 8
Network Topology of a Mobile App
Radio Access Network App Server
Internet
Core Cellular Network
Where is the bottleneck?
3
🤕
SLIDE 9
Network Topology of a Mobile App
Radio Access Network App Server
Internet
Core Cellular Network
Where is the bottleneck?
3
🤕
slow server code?
SLIDE 10
Network Topology of a Mobile App
Radio Access Network App Server
Internet
Core Cellular Network
Where is the bottleneck?
3
🤕
slow server code? throttling at middlebox?
SLIDE 11
Network Topology of a Mobile App
Radio Access Network App Server
Internet
Core Cellular Network
Where is the bottleneck?
3
🤕
congestion? slow server code? throttling at middlebox?
SLIDE 12
Network Topology of a Mobile App
Radio Access Network App Server
Internet
Core Cellular Network
Where is the bottleneck?
3
🤕
congestion? slow server code? throttling at middlebox? slow app code?
SLIDE 13
Network Topology of a Mobile App
Radio Access Network App Server
Internet
Core Cellular Network
Where is the bottleneck?
3
🤕
congestion? slow server code? throttling at middlebox? slow app code?
SLIDE 14
A Developer’s Perspective of Network Bottlenecks
Radio Access Network App Server
Internet
Core Cellular Network
4
SLIDE 15
A Developer’s Perspective of Network Bottlenecks
Radio Access Network App Server
Internet
Core Cellular Network
4
N
- v
i s i b i l i t y
SLIDE 16
The Ideal Metric
Base Station Queues Clients
Radio (wireless) Link
5
SLIDE 17
The Ideal Metric
Base Station Queues Clients
Radio (wireless) Link
5
If empty, traffic bottlenecked
- n the way to the base station.
SLIDE 18
The Ideal Metric
Base Station Queues Clients
Radio (wireless) Link
5
If empty, traffic bottlenecked
- n the way to the base station.
If nonempty, bottleneck is at base station.
SLIDE 19
The Ideal Metric
Base Station Queues
Providers know the status of the queues; but no one else does. BurstTracker estimates this metric at the client.
Clients
Radio (wireless) Link
5
If empty, traffic bottlenecked
- n the way to the base station.
If nonempty, bottleneck is at base station.
SLIDE 20
The Ideal Metric
Base Station Queues
Providers know the status of the queues; but no one else does. BurstTracker estimates this metric at the client.
Clients
Radio (wireless) Link
5
Opportunity: base station scheduling is accessible at the client!
SLIDE 21
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
6
SLIDE 22
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
6
Resources scheduled to purple user in this millisecond.
SLIDE 23
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
7
SLIDE 24
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
7
Base station gives all resources to one user!
SLIDE 25
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
8
SLIDE 26
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
8
SLIDE 27
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
8
Purple user not scheduled due to contention; its queue is still nonempty.
SLIDE 28
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
9
SLIDE 29
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
9
SLIDE 30
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
9
Purple user not given all resources; its queue has drained out.
SLIDE 31
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
10
SLIDE 32
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
11
A user only needs to know their resource allocation to infer their queue status.
SLIDE 33
Resource Allocation Reveals Queue Status
Time (msec) Resource Block (RB) Begin End Burst User
1 Likely 1 2
11
A user only needs to know their resource allocation to infer their queue status.
SLIDE 34
0.00 0.25 0.50 0.75 1.00 50 100
Resources allocated in each time slot Cumulative Probability
Does BurstTracker Generalize?
12
Provider: Verizon
1 Mbps
- ver 1 min
SLIDE 35
0.00 0.25 0.50 0.75 1.00 50 100
Resources allocated in each time slot Cumulative Probability
Does BurstTracker Generalize?
12
Provider: Verizon
1 Mbps
- ver 1 min
SLIDE 36
0.00 0.25 0.50 0.75 1.00 50 100
Resources allocated in each time slot Cumulative Probability
Does BurstTracker Generalize?
13
Slow transfer was aggregated into bursts that used most of the resources.
Provider: Verizon
1 Mbps
- ver 1 min
busy cell idle cell
SLIDE 37
0.00 0.25 0.50 0.75 1.00 50 100
Resources allocated in each time slot Cumulative Probability
Does BurstTracker Generalize?
13
Slow transfer was aggregated into bursts that used most of the resources.
Provider: Verizon
1 Mbps
- ver 1 min
busy cell idle cell
SLIDE 38
0.00 0.25 0.50 0.75 1.00 50 100
Resources allocated in each time slot Cumulative Probability
Does BurstTracker Generalize?
13
Slow transfer was aggregated into bursts that used most of the resources.
Provider: Verizon
1 Mbps
- ver 1 min
busy cell idle cell
SISO MIMO Carrier Agg.
SLIDE 39
Is BurstTracker Accurate?
14
Experiment Setup
- 100 runs of each workload
- Network conditions ~ 2–12.5 Mbps
Partnered with Tier-1 provider to get ground-truth queue status measurements.
Application BurstTracker Median Error (%) File Download 7.2 Video Streaming 6.9
SLIDE 40
Is BurstTracker Accurate?
14
BurstTracker achieves a median error of 7% for different classes of mobile apps.
Experiment Setup
- 100 runs of each workload
- Network conditions ~ 2–12.5 Mbps
Partnered with Tier-1 provider to get ground-truth queue status measurements.
Application BurstTracker Median Error (%) File Download 7.2 Video Streaming 6.9
SLIDE 41
Case Study: Video Streaming
We found that, surprisingly, the LTE downlink was not the bottleneck.
15
- ●
- ●
- 2
4 6 50 100 150 200
time (s) Throughput (Mbps)
speed test average speed test average video streaming
video streaming speed test average speed test before speed test after speed test before video streaming
SLIDE 42
Case Study: Video Streaming
We found that, surprisingly, the LTE downlink was not the bottleneck.
15
- ●
- ●
- 2
4 6 50 100 150 200
time (s) Throughput (Mbps)
speed test average speed test average video streaming
video streaming speed test average speed test before speed test after speed test before video streaming
SLIDE 43
1 Video Segment at the Client
0.00 0.25 0.50 0.75 1.00 200 400 600
time (msec) Fraction of RBs
Burst Start Burst End RBs/msec
16
Resource Allocation Trace for a Single Video Segment
SLIDE 44
1 Video Segment at the Client
0.00 0.25 0.50 0.75 1.00 200 400 600
time (msec) Fraction of RBs
Burst Start Burst End RBs/msec
16
Resource Allocation Trace for a Single Video Segment
SLIDE 45
1 Video Segment at the Client
0.00 0.25 0.50 0.75 1.00 200 400 600
time (msec) Fraction of RBs
Burst Start Burst End RBs/msec
16
BurstTracker indicates that it might be TCP Slow-Start.
Resource Allocation Trace for a Single Video Segment
SLIDE 46
Slow-Start Restart at the Middlebox
25 50 75 100 AT&T T−Mobile Verizon
Link Demand Ratio (%)
HTTP (80) HTTPS (443) unreserved (7777)
- Only HTTP and HTTPS traffic used middlebox.
- Nonstandard port (7777) bypasses middlebox.
17
SLIDE 47
Slow-Start Restart at the Middlebox
25 50 75 100 AT&T T−Mobile Verizon
Link Demand Ratio (%)
HTTP (80) HTTPS (443) unreserved (7777)
18
- Only HTTP and HTTPS traffic used middlebox.
- Nonstandard port (7777) bypasses middlebox.
SLIDE 48
Slow-Start Restart at the Middlebox
25 50 75 100 AT&T T−Mobile Verizon
Link Demand Ratio (%)
HTTP (80) HTTPS (443) unreserved (7777)
19
Split-TCP proxies were forcing Slow-Start restart.
- Only HTTP and HTTPS traffic uses middlebox.
- Nonstandard port (7777) bypasses middlebox.
SLIDE 49
BurstTracker
20
github.com/arjunvb/bursttracker
- Tool to determine if cellular downlink is the bottleneck
- Showed that we can infer base station queue status from resource allocation
- Discovered that carrier’s middlebox was bottleneck for video streaming