• No results found

Distributed Compute for Both Performance and Cost Optimization - Maximizing Data Center Performance in Financial Services

N/A
N/A
Protected

Academic year: 2021

Share "Distributed Compute for Both Performance and Cost Optimization - Maximizing Data Center Performance in Financial Services"

Copied!
25
0
0

Loading.... (view fulltext now)

Full text

(1)

Copyright © 2014 Juniper Networks, Inc. 1

Distributed Compute for Both

Performance and Cost

Optimization

- Maximizing Data Center

Performance in Financial Services

Andrew Bach, Chief Architect for

Financial Services, Juniper Networks

(2)

Copyright © 2014 Juniper Networks, Inc.

2 Copyright © 2015 Juniper Networks, Inc.

Agenda

The Need for High Performance

The Challenge

The Limitations Today

(3)

Copyright © 2014 Juniper Networks, Inc.

3 Copyright © 2015 Juniper Networks, Inc.

Agenda

The Need for High Performance

The Challenge

The Limitations Today

(4)

Copyright © 2014 Juniper Networks, Inc. 4

Timeline of the financial industry

10,000  –  6000  

BCE  

 

Local  Trade  

Market  

OCE  

 

Trade  Routes  

Established  

1604  

 

First  Exchange  

Formed  

1800’s  

Futures  

Market  

Developed  

1800’s  

 

NaIonal    

Markets  

Developed  

1400  

 

First  Banks  

1602  

 

First  Stock  

Issued  

1700’s  

 

REGIONAL    

Exchanges  

Developed  

Late  1900’s  

 

Global  

Exchanges  

Developed  

10,000  –  6000  BCE  

OCE  

1400  

1500  

1600  

1700  

1800  

1900  

(5)

Copyright © 2014 Juniper Networks, Inc. 5

Scale of the business

The total market cap of all listed companies is over

$55 Trillion

….

…Compared to country GDPs, that would make the

sector the largest economy in the world.

 Largest  Economies  by  GDP  (in  Trillions  of  $)  

Germany  -­‐  $3.6T

 

India  -­‐  $1.9T

 

Japan  -­‐  $5.8T

 

China  -­‐  $7.5T  

US  -­‐  $15T

 

$55  Trillion  

Market  Cap

 

(6)

Copyright © 2014 Juniper Networks, Inc. 6

Real money

 DTCC clears, settles over:

§

  $24 quadrillion dollars per year

§

  24 X 10

15

or $24,000,000,000,000,000

§  One US bank alone transacts over $35 Trillion dollars a day

(7)

Copyright © 2014 Juniper Networks, Inc.

7 Copyright © 2015 Juniper Networks, Inc.

400+ Years of Rapid Technology Adoption

And a rich history of technology innovation in markets

• 

First stock ticker to disseminate data (1867)

• 

First telephones on the trading floors (1878)

• 

First electronic ticker display board (1966)

• 

Wireless handheld devices on Trading

floors 15 years before iPad invented (1995)

• 

Industry’s first private network offering

global connectivity

(8)

Copyright © 2014 Juniper Networks, Inc. 8

Transactions become bandwidth

(9)

Copyright © 2014 Juniper Networks, Inc.

9 Copyright © 2015 Juniper Networks, Inc.

Agenda

The Need for High Performance

The Challenge

The Limitations Today

(10)

Copyright © 2014 Juniper Networks, Inc.

10 Copyright © 2015 Juniper Networks, Inc.

• 

Requirement for long

term retention of data

• 

Requirement to archive

meta data

• 

Real time risk

management is now

required

• 

Requirement for

precision timing

• 

Increasing focus on

cyber security

• 

Time to market must be

faster

• 

Product life time is

shorting

• 

Margins are driving OPEX

reduction

• 

Regulation is increasing

the cost of business

• 

Technology is a

strategic weapon

• 

Fueling the race to the

triple crown of

technology (0 cost, 0

latency, 0 time to

market)

• 

Bandwidth demand

continues to grow at

30% - 50% per year

• 

Rate of data growth is

also increasing

• 

Exchange trading floors

continue to fade away

• 

Machine assisted

trading dominates

• 

Heuristic trading based

on hyper-contextual

information is growing

• 

Risk assessment is now

routinely based on

massive near real time

data bases

• 

Shared public/private

cloud solutions

Regulatory Model

is Driving Change

FSS is Evolving to

a Commodity

Industry

Technology Adoption

is Accelerating to

Meet Accelerating

Business Needs

Business Model

Continues to Evolve

(11)

Copyright © 2014 Juniper Networks, Inc. 11

Customer

A Typical Challenge – Exchange Gateway Trade

Plant

Total Latency

≈150.0µS

Trade Plant Size:

≈ 100 Servers

≈ 1500 Ports

(12)

Copyright © 2014 Juniper Networks, Inc. 12

A Typical Challenge – Market Data Ticker Plant

Internal

Users

Feed handler &

publishing servers

server farm

Pub/sub

Client subscriber

servers

World Wide

Market Data

(13)

Copyright © 2014 Juniper Networks, Inc. 13

Next Challenge – Heuristic Based Trading

Market data

feeds

News feeds

Social feeds

Streaming

news video

Ticker Plant

Heuristic

Processing

Heuristic

Processing

Heuristic

Processing

Heuristic

Processing

Historical

Data Base

Heuristic

Processing

Human

Oversight

and control

Execution

Venues

Trade Plant

(14)

Copyright © 2014 Juniper Networks, Inc.

14 Copyright © 2015 Juniper Networks, Inc.

Agenda

The Need for High Performance

The Challenge

The Limitations Today

(15)

Copyright © 2014 Juniper Networks, Inc. 15

Slowing of Processor Speed

(16)

Copyright © 2014 Juniper Networks, Inc. 16

Latency Reduction Trends

(17)

Copyright © 2014 Juniper Networks, Inc.

17 Copyright © 2015 Juniper Networks, Inc.

Agenda

The Need for High Performance

The Challenge

The Limitations Today

(18)

Copyright © 2014 Juniper Networks, Inc.

18 Copyright © 2015 Juniper Networks, Inc.

At about 200 – 500NS for

a reasonable switch

Need to focus on a

different approach

Lower latency

Eliminate servers

Reduce network ports

Application can be

embedded in

A VM in the switch

A FPGA in the data

path of the switch

ASIC Packet

forwarding engine

(PFE)

The Race to Zero is

Ending

Embed Application

Snippets into the

Switching Fabric

Embed Snippets at the

Control or Data Plane of

the Network

Compute-Integrated Networking

(19)

Copyright © 2014 Juniper Networks, Inc.

19 Copyright © 2015 Juniper Networks, Inc.

What Types of Computing Can We Use

CPU

• 

Good for general purpose and data flow management

FPGA

• 

Highly optimizable field programmable hardware executing

custom code and data flow tasks

ASIC

• 

Highly optimized purpose built processor, a Packet Forwarding

Engine (PFE) for example

• 

Merchant silicon

(20)

Copyright © 2014 Juniper Networks, Inc. 20

Network Switch/Application Acceleration/Packet Forwarding Accelerator

PFE  

Intel  X86  ,  DDR3,  SSD  

 

HYPERVISOR  

NETWORK  OS  

ACTIVE  

NETWORK  OS    

STANDBY  

GUEST  APPLICATION(S)  

(21)

Copyright © 2014 Juniper Networks, Inc.

21 Copyright © 2015 Juniper Networks, Inc.

Programing distributed compute systems cost effectively

• 

For X86 environment with VM/Containers either HLL or third party

libraries

• 

For FPGA

Data Flow

OpenCL

(22)

Copyright © 2014 Juniper Networks, Inc. 22

Customer

Risk

Management

embedded in

Data center

Switch

Symbol

routing

embedded in

Data center

Switch

Data pre/post

processing

Customer

40% less

ports Latency

down 50µS

Total Latency

≈100.0µS

Trade Plant Size:

≈ 60 Servers

≈ 1000 Ports

Total Latency

≈150.0µS

Trade Plant Size:

≈ 100 Servers

≈ 1500 Ports

A Different Approach: Distributed Computing for

an Exchange Gateway Trade Plant

(23)

Copyright © 2014 Juniper Networks, Inc. 23

A Different Approach: Market Data Ticker Plant

Internal

Users

Feed handler &

publishing servers

server farm

Pub/sub

Client subscriber

servers

World Wide

Market Data

Feeds

Internal

Users

World Wide

Market Data

Feeds

Feed handler &

publishing in top of

rack switch

Pub/sub

server farm

Client subscriber

imbedded in top of

rack switch

(24)

Copyright © 2014 Juniper Networks, Inc.

24 Copyright © 2015 Juniper Networks, Inc.

Processors and network switches are

hitting natural limits

To achieve a high performance infrastructure compute resource must be

distributed

Optimize computing in the server, NIC (FPGA), and data flow engines

Architect compute resources across the data center - not just one place

Centralize Processing Where You Can, Distribute

Where You Must

(25)

References

Related documents

The conceptual framework for this research study included six areas that influenced student success in preparation for college mathematics: improving communication and aligning

The last chapter includes the main result, this is the distributional limit the- orem for linear fractional stable motions including a representation of a linear fractional

After briefly exposing salient features of VARs (paragraph 1), a first set of solutions active on sample size is investigated: paragraph 3 outlines major temporal

Needed artifacts: Architectural Diagrams, Use Case Scenarios and Traces, Misuse Case Scenarios and Traces, Attack Trees, Essential Services and Asset Identification,

The intangible value of the business, above and beyond the fixed, tangible assets must be allocated to specific intangible asset categories and are amortized (definite lived) or

Persistent does have a defined Knowledge Management sytem but there are efforts taken where the features of knowledge management are implemented like company’s

To convert concentrations in mg/L (or ppm in dilute solution) to molarity, divide by the molar mass of the analyte to convert mass in mg into a corresponding number of moles.. What