Riptide: Jump Starting Back-Office Connections in Cloud Systems - - PowerPoint PPT Presentation

▶

Mar 06, 2023 148 likes •633 views

Riptide: Jump Starting Back-Office Connections in Cloud Systems Marcel Flores - Northwestern University Amir R. Khakpour - Verizon Digital Media Services Harkeerat Bedi - Verizon Digital Media Services 1 Cloud systems Large scale global

SLIDE 1

Riptide: Jump Starting Back-Office Connections in Cloud Systems

Marcel Flores - Northwestern University Amir R. Khakpour - Verizon Digital Media Services Harkeerat Bedi - Verizon Digital Media Services

SLIDE 2

Cloud systems

Large scale global services:
CDNs, web services.
Back-office traffic between Points of Presence (PoPs).
Control messages, small transfers.

SLIDE 3

Cloud systems

Frequent opening of

connections between PoPs.

In a perfect world,

would have a mesh.

Application and

resource constraints limit reuse.

SLIDE 4

Cloud systems

Frequent opening of

connections between PoPs.

In a perfect world,

would have a mesh.

Application and

resource constraints limit reuse.

SLIDE 5

Cloud systems

Frequent opening of

connections between PoPs.

In a perfect world,

would have a mesh.

Application and

resource constraints limit reuse.

SLIDE 6

Cloud systems

Frequent opening of

connections between PoPs.

In a perfect world,

would have a mesh.

Application and

resource constraints limit reuse.

SLIDE 7

Cloud systems

Frequent opening of

connections between PoPs.

In a perfect world,

would have a mesh.

Application and

resource constraints limit reuse.

SLIDE 8

Cloud systems

Frequent opening of

connections between PoPs.

In a perfect world,

would have a mesh.

Application and

resource constraints limit reuse.

SLIDE 9

Slow-start penalty

Sender Receiver

SLIDE 10

Slow-start penalty

Sender Receiver Syn

SLIDE 11

Slow-start penalty

Sender Receiver Syn Syn/Ack

SLIDE 12

Slow-start penalty

Sender Receiver Syn Syn/Ack

{

Initial Window

SLIDE 13

Slow-start penalty

Sender Receiver Syn Syn/Ack

{

Initial Window

Data doesn’t fit

SLIDE 14

Slow-start penalty

Sender Receiver Syn Syn/Ack

{

Initial Window

{

2nd Window

Data doesn’t fit

SLIDE 15

Slow-start penalty

Sender Receiver Syn Syn/Ack

{

Initial Window

{

2nd Window

Data doesn’t fit Transfer pays second RTT

SLIDE 16

Cloud workloads

200 400 600 800 1000 File Size (KB) 0.0 0.2 0.4 0.6 0.8 1.0 CDF

SLIDE 17

Cloud workloads

200 400 600 800 1000 File Size (KB) 0.0 0.2 0.4 0.6 0.8 1.0 CDF

54% are too big for default

SLIDE 18

Cloud workloads

200 400 600 800 1000 File Size (KB) 0.0 0.2 0.4 0.6 0.8 1.0 CDF

54% are too big for default

85% under 200KB

SLIDE 19

Global deployments

50 100 150 200 250 300 350 400 RTT (MS) 0.0 0.2 0.4 0.6 0.8 1.0 CDF

SLIDE 20

Global deployments

50 100 150 200 250 300 350 400 RTT (MS) 0.0 0.2 0.4 0.6 0.8 1.0 CDF

Median RTT is over 125 ms

SLIDE 21

Global deployments

Can’t just blindly increase the congestion window
n a global deployment.
Would risk significant loss.

SLIDE 22

Riptide

Observes current congestion windows.
New connections set initial window to a known-safe level.
Operates in a totally standalone manner.

SLIDE 23

Riptide

Riptide observes CWND

for all open connections to a destination.

New connections will be
pened with INIT_CWND

set to the average of existing windows.

Riptide

85 75

SLIDE 24

Riptide

Riptide observes CWND

for all open connections to a destination.

New connections will be
pened with INIT_CWND

set to the average of existing windows.

Riptide

85 75 ?

SLIDE 25

Riptide

Riptide observes CWND

for all open connections to a destination.

New connections will be
pened with INIT_CWND

set to the average of existing windows.

Riptide

85 75 80

SLIDE 26

Implementation

Implemented as a Python script in user space.
Use the ss tool to observe existing windows.
Polls current connections once per second.
Sets new windows via ip route interface.

ip route add 10.0.0.127 dev eth0 proto \\ static initcwnd 80 via 10.0.0.1

SLIDE 27

Riptide Deployment

SLIDE 28

Probes

To test the current state of the network, send hourly

probes between PoPs.

Currently employ 10K, 50K, 100K probes.

SLIDE 29

Observed windows

SLIDE 30

Observed windows

CWND windows significantly higher.

SLIDE 31

Observed windows

CWND windows significantly higher.

SLIDE 32

Observed windows

CWND windows significantly higher. Knee

SLIDE 33

Probe completion times

Riptide Default

Clients are able to

complete the probe transfers in fewer round trips.

Reduces total transfer

time.

SLIDE 34

Probe completion times

Riptide Default

Clients are able to

complete the probe transfers in fewer round trips.

Reduces total transfer

time. Better

SLIDE 35

Probe completion times

Riptide Default

Clients are able to

complete the probe transfers in fewer round trips.

Reduces total transfer

time. Better

SLIDE 36

Probe completion times

Riptide Default

Clients are able to

complete the probe transfers in fewer round trips.

Reduces total transfer

time. 80% of transfers were faster Better

SLIDE 37

Gain Percentile

SLIDE 38

Gain Percentile

}

SLIDE 39

Gain Percentile

Gains were highest at upper percentiles.

}

SLIDE 40

Conclusion

Demonstrated design and implementation of a

simple tool to observe and adjust congestion windows.

Deployed the system in a production network.
Achieved significant increase in average

congestion window.

Demonstrated improvements in completion time,

reducing slow-start penalty

SLIDE 41

Thank you!

SLIDE 42

Extras

SLIDE 43

Cloud systems

Complexity means node-level resource constraints
Frequent connections between Points-of-Presence

(PoPs).

In many cases dominated by small file

transactions.

SLIDE 44

Cloud workloads

1 2 3 4 5 6 7 8 Number of RTT s 0.0 0.2 0.4 0.6 0.8 1.0 CDF

10 25 50 100

SLIDE 45

Cloud workloads

1 2 3 4 5 6 7 8 Number of RTT s 0.0 0.2 0.4 0.6 0.8 1.0 CDF

10 25 50 100

Larger windows reduce RTTs

SLIDE 46

Traffic matters

SLIDE 47

Traffic matters

Traffic drives up window sizes.