Muhammad Shahbaz Tushar Swamy Alex Rucker Neeraja Yadwadkar Yaqi Zhang and Kunle Olukotun Taurus An Intelligent Data Plane Programmable Data Plane fused with Machine Intelligence ID: 814923
Download The PPT/PDF document "Taurus: An Intelligent Data Plane" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Taurus: An Intelligent Data Plane
Muhammad ShahbazTushar Swamy, Alex Rucker, Neeraja Yadwadkar, Yaqi Zhang, and Kunle Olukotun
Slide2Taurus: An Intelligent Data Plane
Programmable Data Planefused withMachine Intelligence
Deep Learning
Slide3Managing Networks is Hard!
Cloud Computing
Internet of Things (IoT)
Augmented and Virtual
Reality (AR/VR)
Slide4Approaches to Manage Networks are …
Data Plane
Switch or NIC
Fast
yet
dumb
Slow
but
intelligent
Control Plane
Intelligence
Slide5Approaches to Manage Networks are …
Data Plane
Switch or NIC
Fast
yet
dumb
Examples:
Congestion control
Load balancing (ECMP, RSS)
Queue scheduling
and more
Characteristics:
Operates on packets or
flowlets
(
i.e.,
bursts of packets)
Uses heuristics … hash, etc.
Low latency … ≤ sub µs
High throughput …
Tbps
Packets In
Packets Out
Slide6Approaches to Manage Networks are …
Data Plane
Switch or NIC
Fast
yet
dumb
Slow
but
intelligent
Control Plane
Intelligence
Slide7Approaches to Manage Networks are …
Data Plane
Switch or NIC
Slow
but
intelligent
Control Plane
Intelligence
Examples:
Anomaly detection
Automation
Recommendation
Characteristics:
Operates on flows
Performs complicated tasks
Sub-second latency
Low throughput
Packets In
Packets Out
Packet
Digest
Flow
Rule
Slide8Approaches to Manage Networks are …
Data PlaneSwitch or NICFast yet dumb
Slow but intelligent
Control Plane
Intelligence
Slide9Approaches to Manage Networks are …
Data PlaneSwitch or NICFast yet dumb
Slow but intelligent
Control Plane
Intelligence
Slide10Approaches to Manage Networks are …
Data Plane Switch or NICFast and
Control Plane
Intelligence
Intelligence
intelligent
intelligent
Slide11Data Plane
Switch or NICTaurus: An Intelligent Data PlaneIntelligence
Packets In
Packets Out
Slide12Taurus: An Intelligent Data Plane
Packet
Parser
Match-Action
Tables
Map Reduce
Unit
Match-Action
Tables
Traffic
Manager
Packets In
Packets Out
Slide13Traffic
ManagerMatch-ActionTablesMap ReduceUnitMatch-ActionTables
Taurus: An Intelligent Data Plane
Packets In
Packets Out
Packet
Parser
Implements a
finite state machine (FSM)
that operates on a user-defined
parse graph
Converts the
incoming packet bit stream into vectors
,
e.g.
,
headers (IP or TCP)
features (switch, host, or other encoded state)
Slide14Traffic
ManagerTaurus: An Intelligent Data Plane
Packet
Parser
Map Reduce
Unit
Packets In
Packets Out
Match-Action
Tables
Match-Action
Tables
Memory
ALU
A
match-action table:
Memory for
exact
(SRAM) and
ternary
(TCAM) match
ALU for basic single-cycle
VLIW operations
(no loops or multiplication)
Performs pre- and post-processing, and switching tasks:
Pre-processing
,
e.g.
, data transformation, acquisition, and repair
Post-processing
,
e.g.
, drop, route, or encode output of the learned model on a packet
Slide15Traffic
ManagerTaurus: An Intelligent Data Plane
Packet
Parser
Match-Action
Tables
Match-Action
Tables
Packets In
Packets Out
Map Reduce
Unit
Implements a
spatial SIMD architecture
, composed of
compute
and
memory
units
Compute consists of
pipelined
stages of
functional units
(FUs) and
pipeline registers
(PRs), running in
parallel
as lanes
Processes parallel
map-reduce operations
to accelerate
linear algebra
, crucial for
machine learning (ML)
, or other network functions
Slide16Taurus: An Intelligent Data Plane
Packet
Parser
Match-Action
Tables
Map Reduce
Unit
Match-Action
Tables
Packets In
Packets Out
Traffic
Manager
Responsible for
storing
and
forwarding
packets off of the chip:
Queuing
: buffer incoming packet
Replication
: clone packets across multiple egress ports (
e.g.
, multicast)
Scheduling
: forward packets based on a queuing discipline (
e.g.
, PIFO) or instructions from the match-action tables
Slide17Taurus: An Intelligent Data Plane
Packet
Parser
Match-Action
Tables
Map Reduce
Unit
Match-Action
Tables
Packets In
Packets Out
Traffic
Manager
Slide18Example: Anomaly Detection
Packet
Parser
Match-Action
Tables
Map Reduce
Unit
Match-Action
Tables
Packets In
Packets Out
Parse incoming
packets
and
read local features
like an IP address
Retrieve
out of
network events
like no. of failed
logins for a
connection
Process features
using a learned
function
and output probability
of an anomalous packet
Parsing
Pre-processing
using tables
Machine
learning
Select a port
based on the
probability
of an anomalous packet
Post-processing
using tables
Send packets
out over the
selected port
Scheduling
Traffic
Manager
Slide19Evaluation
[4] Design principles for packet parsers, ACM/IEEE ANCS’13[5] CONGA: distributed congestion-aware load balancing for datacenters, ACM SIGCOMM’14[6] Packet Transactions: High-Level Programming for Line-Rate Switches, ACM SIGCOMM’16
10
6
x
Within
2 – 15%
[5,6]
Slide20Future Work
Packet
Parser
Match-Action
Tables
Map Reduce
Unit
Match-Action
Tables
Packets In
Packets Out
Traffic
Manager
Spatial
Parse
Graph
MAT
MAT
PIFO
Ad-hoc Low-Level DSLs
A Unified High-Level Abstraction
?
Conclusion
Designed to run machine-learning inference inside a data planeProvides orders of magnitude improvement over existing approaches
Data Plane
Switch or NIC
Fast
and
Intelligence
intelligent
Taurus
Slide22Muhammad Shahbaz
http://cs.stanford.edu/~mshahbaz Questions?