Interview talk at various universities and labs

1
Flowlet Switching
Srikanth Kandula Shan Sinha Dina Katabi
2
ISPs Want to Split Traffic Across Multiple Paths
3
ISPs Want to Split Traffic Across Multiple Paths
70
30

• Load balancing to remove hot spots
• Rebalance traffic when unpredictable events occur
  (Outages, DoS, BGP reroutes, Flash Crowds, )

4
ISPs Want to Split Traffic Across Multiple Paths
Unpredictable Traffic
Rebalance Traffic
70
30

• Load balancing to remove hot spots
• Rebalance traffic when unpredictable events occur
  (Outages, DoS, BGP reroutes, Flash Crowds, )

5
ISPs Want to Split Traffic Across Multiple Paths
Unpredictable Traffic
30
70

• Load balancing to remove hot spots
• Rebalance traffic when unpredictable events occur
  (Outages, DoS, BGP reroutes, Flash Crowds, )

6

• Much research on balancing and rebalancing load,
• But implementation is hard particularly with
  dynamic ratios
• Either sacrifice accuracy or reorder TCP packets

7

• Much research on balancing and rebalancing load,
• But implementation is hard particularly with
  dynamic ratios
• Either sacrifice accuracy or reorder TCP packets

Problem

1. Given the desired split ratios possibly dynamic
2. Split traffic accurately, at the edge router,
   without reordering TCPs packets

8
Existing Scheme 1 Packet-Based Splitting

• Assign packets to paths proportional to the
  desired ratios
• Reorders TCP packets causing bad throughput

9
Existing Scheme 2 Flow-Based Splitting

• Assign TCP flows to each path proportional to the
  desired ratio
• Flows are not all equal Elephants Mice
• So, estimate the rate of each TCP flow
• But rates change with time
• Too complex
• Very inaccurate if desired ratios change

10
How to Split Traffic?

• Packet-Based
• Accurate
• Reorders TCP packets
• Easily tracks dynamic ratios

• Flow-Based
• Inaccurate
• No packet reordering
• Hard to track if ratios change

Can we combine the best of the two approaches?
11
This Talk

• Show how to send a single TCP flow down multiple
  paths without reordering
• Accurately split traffic even when desired ratios
  are dynamic
• Easy to implement

12
Flowlet Switching
1
2

• If the previous packet from the flow has left the
  merging point ? Can reassign the flow to a
  different path

13
Flowlet Switching
Given ? gt D2-D1
14
Flowlet Switching
Given ? gt D2-D1
Flowlets are bursts from same flow separated by
at least ? they can be switched independently!
Idle ?
15
Implementing Flowlet Switching is Simple
hash
SRCip DSTip SRCPort DSTPort

• Router at the split point hashes packet header
• If (Now - Last_Seen) gt ?, flow can change path
• Reassign path proportionally to the desired split
  ratios

16
Does it Really Work?

• Traces collected on a peering link, an edge link
  and two core links
• Split Vectors (3 paths)
• Static (.3, .3, .4)
• Dynamic sinusoidal with amplitude 60, period
  20min Akella04,Chuah02

17
Is Flowlet Switching Accurate?
Error
18
Is Flowlet Switching Accurate?
Error
Flowlet switching is much more accurate than
flow-based switching
19
Can do Flowlet Switching without Per-Flow State
Fig. shows Avg. and Max. of many traces
4 16 64 256 1024
2048 4096 8192

Hash Table Entries
Active Flows 50,000 But Router maintains a
hash table lt 1000 entries (5KB).
20
Understanding Flowlets
21
But Where do Flowlets come from?

• Cant be just timeouts or short flows most of
  the bytes are in the elephants
• Why can a large flow be broken into many small
  flowlets?

22
Flowlets exist because TCP is bursty at RTT and
sub-RTT scales

• Well-known that TCP usually sends a window in one
  or a few bursts and waits for acks
  Zhang91,Zhang03, Jiang04
• Some Reasons
• Slow-start
• Ack compression
• Window is much smaller than delay-BW product

23
Flowlets exist because TCP is Bursty
Most flowlets have inter-arrivals less than an
RTT ? most flowlets are sub-windows
24
Why Flowlet Switching is Accurate?

• 80 of bytes are in flowlets smaller than 10KB
• Assigning a flowlet to a path isnt a long
  commitment

25
Why Flowlets can Track Dynamics?
Arrival Rate of both flows and flowlets (/sec)
143.16
1454.98
Edge
611.95
8661.43
Peering
3784.10
35287.04
Core1
111.33
2848.76
Core2
An order of magnitude more opportunities to
rebalance!
26
Why flowlet switching doesnt need per-flow state?
27
Why flowlet switching doesnt need per-flow state?
Flow 1
Flow 2
Flow 3
28
Why flowlet switching doesnt need per-flow state?
Flow 1
Flow 2
Flow 3
29
Why flowlet switching doesnt need per-flow state?
Flow 1
Flow 2
Flow 3
Active Flowlets
Time
30
Why flowlet switching doesnt need per-flow state?
Trace
Edge
Peering
Core1
Core2
31
Why flowlet switching doesnt need per-flow state?
Trace
Edge
Peering
Core1
Core2
Active flowlets is 2 orders of magnitude smaller
than flows ? Very small hash table
32
Why Flowlet Switching is Possible?

• Why can a large flow be broken into many small
  flowlets?
• Why is flowlet switching accurate?
• Why flowlet switching does not need per-flow
  state?

• TCP burstiness at small time scales
• Small commitment many more chances to rebalance
• Few simultaneously active flowlets

33
Configuring Flowlet Switching
Flowlet separation gt delay difference
But, how to find delay difference?

• For our traces which are a diverse collection of
  traffic within continental US
• 50ms is a good and safe choice!
• Our procedure is a constructive way to find ?

34
Flowlet Separation of 50ms is Good
Any flowlet timeout in 50, 100 ms yields highly
accurate splits
35
Flowlet Separation of 50ms is Safe
1 .8 .6 .4 .2 0
Even if delay difference gtgt 50ms, prob. of
reordering is negligible compared to drop. rate
in the Internet (about 1)
36
Conclusion

• Harness TCP burstiness to split traffic at a
  finer resolution than a flow without reordering
• Flowlet Switching
• Splitting errors are a few percents
• Reordering probability is negligible compared to
  drop prob. in the Internet
• Easy to implement
• Enable ISPs to do dynamic load balancing

37
More Information athttp//nms.lcs.mit.edu/dina/
texcp.html
Questions?
38
For 50ms, very few retransmissions triggered even
when delay difference is severely under-estimated
Actual Delay Diff Flowlet Timeout
Flowlet Timeout
39
Flowlet Switching has Negligible Overhead
Error
4 16 64 256
1024 2048 4096 8192

Hash Table Size
Very, Very Cheap. Edge routers maintain a hash
table of 210 entries (lt10KB).
40
Any flowlet timeout in 50, 100 ms yields highly
accurate splits
Error
Flowlet Separation (msec)
41
Flowlet Switching Tracks Dynamic Splits
For clarity, we show only one path!