第一周

Introduction

adaptive streaming analytics need both data fidelity and data freshness

TCP ensures data fidelity but not data freshness

UDP ensures data freshness but not data fidelity

Manual policies are sub-optimal, cannot generalize

So solve the problem in SYSTEM-level

• Abstraction
• Learn policy other than use manual one
• Runtime system

RESULTS

• 40x faster than TCP
• 45% imporvment than UDP
• Better than JetStream

Motivation

Use public Internet reduce cost but show congestions

Real-time monitoring is very high-demending

WAN resources' scarcity

Awstream design

API

Adaptation as first-class abstraction

Automatic Profiling

$c = [k_1,k_2,\dots,k_n]$

Calculate $B(c_i)$[bandwidth requirement for c] and $A(c_i)$[accuracy measure for c]

Find Pareto-optimal set $P$, where

$P = \{c \in C:{c'\in C:B(c')<B(c),A(c')>A(c)} = \varnothing\}$

using offline profiling or online profiling

Runtime Adaptation

Resource Allocation & Fairness

max the mininal acc in all applications（木桶原理）

Implementation

Evaluation

large bandwidth variation

single knob(one dimension) leads to sub-optimal results

Runtime adaptation

use Linux tc(use man tc) to control the clients’ outgoing bandwidth

Background Limitation is 25Mbps
t < 200s no shaping 
200s <= t <= 380s limit to 7.5Mbps
380s < t <= 440s limit to 5Mbps

Discussion and future work

reduce dev effort

fault tolerance and recovery

profile modeling

bandwidth modeling and prediction

Related work

/