Transcript Document

The Next Phase of the
Data Center Evolution:
Heterogeneous Computing
Bob Doud, Director of Marketing
Tilera Corporation
1
Changes Are Coming to the Data Center
• Datacenters are experiencing an enormous increase in
traffic, as well as the drive for flexibility and scale
• As a result, operators need more:
•
•
•
Agility
Power Efficiency
Performance Density
• New technologies such as
SDN and NFV are ushering in
a new paradigm
2
Network Functions Migrating to Commodity Servers
Proprietary hardware
Server-based hardware
• Long development and qual. cycles
• Stifled innovation due to fixed HW
- memories, CPUs, I/Os, etc.
• Limited economy of scale (low volumes)
• Flexibility & elasticity
• Readily available from many ODMs
• Massive economy of scale (high volumes)
3
26,000
But New Workloads Create a Challenge
Workloads x VMs x Traffic >>> Compute
18,000
9500
8500
6000
CPU
cycles/packet
5000
2250
1500
70
175
750
4
The Growing Gap Between Compute and Networking
Growing gap between
server compute and
networking compute
GIGAOM JUN. 19, 2013 - 12:15 PM PDT
Facebook experiences 1,000
times more traffic inside its
data centers than it sends and
receives from outside users.
That traffic growth has
promoted the social network to
design a fabric networking
architecture.
5
Server Processors Aren’t Architected for Dataplane
Server CPU
User-space applications are starved
Hypervisor
30%
VM / Applications
33%
Table Search
15%
22%
N-Tuple Classify
Hypervisor consumes majority of CPU
~ 30%
Application-centric processing
~ 70%
Data Plane-centric processing
Flow Rx/Tx
Server Processor
workload profile
6
One Solution: Buy More Servers
And more ToR switches
And more floor space
And more power
And more cooling
And…
Google Iowa data center
7
Or… Embrace Heterogeneous Computing
• Apply the proper type of compute to match the workload
•
•
x86 CPU: Good for single-threaded, high-touch workloads
Dataplane Processor: Optimum for packet / flow processing,
security processing, deep packet inspection, etc.
+
x86 Server
Intelligent Adapter
• This is not a new concept… think video accelerators!
8
Improve Performance, Density, Power Efficiency
~400W
~350W
+
x86
Server
30%
VM / Applications
33%
Table Search
15%
N-Tuple Classify
22%
Flow Rx/Tx
Control
Plane
Intelligent
Adapter
Control
Plane
Data
Plane
Data
Plane
Networking Workload Profile
VM / Applications
82%
I/O DMA
VMM
Flow Rx/Tx
6%
7%
5%
9
Still Looking for the Right Stuff?
•
We’ve had NIC and accelerator cards for years…
•
But they only tackle part of the problem:
• Limited function cards; most cannot run open source code
• Difficult, non-scaling programming model
•
•
Non-coherent memory (can it run Linux?)
Proprietary tools (i.e. microcode, NPU coding, custom language/tools)
• I/O and PCIe limitations – throughput and/or latency
10
What Embodies the Ideal Solution?
• Intelligence…
•
•
•
•
Fully “Software-Defined” offload, not hardened feature set
Adaptability to changing datacenter needs
Feature Velocity: Ability to deploy new services quickly
Standard programming: Linux, C, GNU/gcc environment
11
What Embodies the Ideal Solution?
• Intelligence…
• Performance…
•
•
•
Wire-speed I/O across packet sizes
Multiple 10G/40G ports, wide PCIe to host CPU
Horsepower equivalent to at least 1 socket of x86
12
What Embodies the Ideal Solution?
• Intelligence…
• Performance…
• Low Power and Size
•
•
Fit the physical and thermal envelope of a PCIe slot
Increase the overall efficiency of the server system
13
COTS x86
The New Heterogeneous NFV Architecture
Control (Application) Plane Scales
Hypervisor cycles are scaled down
Intelligent Adapter
Hypervisor
PCIe
VF
VF
VF
VF
VF
VF
VF
VF
Open
vSwitch
VF
SR-IOV
Data Plane Scales
10 / 40 Gbps Interfaces
14
Open vSwitch (OVS) is a Representative Use Case
•
OVS is an open source multi-layer soft switch implementation
designed for virtualized environments, offering:
•
•
•
•
Flow lookups and load-balancing across VM’s
Protocol tunneling (GRE, VXLAN, etc.)
QoS and traffic policing
VM state mobility (migration)
•
And it can be managed via OpenFlow
•
But, there are drawbacks to running on the x86…
http://openvswitch.org/
15
Host Side OVS is Inefficient
•
•
•
Breaks some of the offload capabilities in most NIC cards
VxLAN encapsulation breaks TSO/LRO in current NIC
implementations
Open vSwitch in hypervisor typically breaks SR-IOV between
NIC and host VM’s
• Must route all traffic through HV and use expensive memcopy
•
Need to over-provision vSwitch for worst-case bandwidth
•
These limitations are removed if OVS is run on the PCIe adapter
• Can deliver packets direct to user-space with DPDK Poll-Mode Driver
16
COTS x86
Enabling Sophisticated Services on the NIC
x86 CPU Dedicated to delivering
applications, not networking
Hypervisor
VF
DPI
VF
IPsec
VF
SSL
VF
IDS/IPS
VF
VF
TCP/IP
VF
L4-7 Load
Balance
VF
Monitoring
VF
Tunneling
Intelligent Adapter
PCIe
Wire-Speed networking
and security services
Open
vSwitch
10 / 40 Gbps Interfaces
17
Multiple Data Center Applications can be Offloaded
Switching, Tunneling, etc.
(10 – 80 Gbps)
NFV,
Open vSwitch
Offload Crypto &
Packet Processing
(10 – 40 Gbps)
IPsec / SSL
Termination
Packet Pre-Processing,
Header Manipulation,
Security Processing
(20 – 40 Gbps)
Application
Delivery
Controller
Network
Monitoring &
Brokering
IDS / IPS
Anti-DDoS
Parse, Classify &
Capture Traffic
(10, 20, 40 Gbps)
Deep Packet
Inspection
(10 – 40 Gbps)
18
Network Analytics/Monitoring
Big Data Analytics
Hypervisor
40Gbps L4-7 DPI
N-Tuple Linux
Classifier/Action
VF
VF
VF
DPI
VF
IPsec
VF
SSL
VF
IDS/IPS
VF
TCP/IP
VF
L4-7 Load
Balance
VF
Monitorin
Intelligent Adapter
PCIe
Open
vSwitch
1-8 10Gbps Interfaces
19
L4 – L7 Load Balancing
L4-7 Balancing
Logic
Hypervisor
40Gbps SSL/TCP
Termination
N-Tuple Linux
Classifier/Action
VF
VF
VF
DPI
VF
IPsec
VF
SSL
VF
IDS/IPS
VF
TCP/IP
VF
L4-7 Load
Balance
VF
Monitorin
Intelligent Adapter
PCIe
Open
vSwitch
1-8 10Gbps Interfaces
20
Network Monitoring/Recording (Cyber Security)
Recording and
offline/online Forensics
Hypervisor
80Gbps L2-4
Classification & Filtering
N-Tuple Linux
Classifier/Action
VF
VF
VF
DPI
VF
IPsec
VF
SSL
VF
IDS/IPS
VF
TCP/IP
VF
L4-7 Load
Balance
VF
Monitorng
Intelligent Adapter
PCIe
Open
vSwitch
1-8 10Gbps Interfaces
21
Network Functions Offload with SDN Management
Server CPU
Guest Guest Guest Guest
OS
OS
OS
OS
VMM
Openflow
PCIe
VF
VF
VF
VF
L2-L7 Linux
Classifier
10G/40G Ethernet
DATAPLANE OFFLOAD
Functions controlled in the NIC:
1) Flow assignments to queues
2) Bandwidth management per flow
3) VM to flows assignments
4) Security policies
5) Metering and Monitoring
6) Local switching and routing
7) Metadata extraction
8) Future: iSCSI, tunneling, etc.
22
Summary
• Datacenter growth and networking workloads
such as NFV are outpacing server compute
• Heterogeneous computing brings the right type
of processing to suit the workload, improving:
•
•
•
Performance
Power Efficiency
Real Estate
23
Thank You
Bob Doud, Tilera corp.
[email protected]
24