Industrial Issues in Software Performance

1
Industrial Issues in Software Performance

• Aman Shaikh
• ATT Labs - Research
• ashaikh_at_research.att.com
• June 24, 2008
• 7th International Workshop on Software and
  Performance
• Princeton

Acknowledgement Alan Glasser, Kobus van der
Merwe, Jennifer Yates
2
Introduction

• Communication service providers
• Various services voice, video, data
• Various ways of connecting to customer
• Ethernet, Frame Relay, ATM, Wireless,
• Trend towards supporting all communication over
  IP-based core network
• Todays IP networks and services
• Thousands of equipments and systems have to work
  in tandem
• Routers, switches, servers, line cards
• Provisioning, configuration, trouble-shooting,
  maintenance, capacity planning, billing,
• Huge amalgamation of hardware and software

3
Challenges of Managing Networks

• Lack of management support
• Absence of standardization in network management
• Solutions often develop in an organic fashion
• Scale
• Hundreds of offices, thousands of routers,
  hundred thousands of line cards and transport
  facilities
• Heterogeneity
• Different kinds of equipments
• Example routers, switches, servers, line cards,
  
• Different vendors and models for same equipment
• Example of router vendors Cisco, Juniper,
  Alcatel,
• Cisco router models CRS, GSR, 7200 series,
• Variety of service offerings with wide variety of
  features

4
Challenges of Managing Networks

• Constant churn
• Customers, services, technology
• Explosive growth
• Internet traffic is predicted to double every two
  years
• http//newsroom.cisco.com/dlls/2008/ekits/Cisco_Vi
  sual_Networking_Index_061608.pdf
• More stringent requirements and expectations
• Service delivery model of IP is best effort
• Works for original applications of the Internet
  file transfer, e-mail and newsgroup
• Applications such as VoIP, IPTV, gaming require
  strong performance guarantees from the network
• Deliver packets reliably within a certain time
  bound
• Security
• Internets intrinsic model Trust every packet
  and do not charge for it

5
Tier-1 Service Provider Network
DWDM systems
OC-48 or OC-192 DWDM
PoP
Intercity
PoP
Customer facing PE interfaces
PoP
Metro

• Rough Numbers
• 100s of offices
• 100s of Ps, 1000s of PEs, 10000s of CEs
• 100,000s of transport facilities

LEC
CE
CPE
Access
CE
CPE
PoP Point-of-Presence P Backbone (core)
Router PE Provider Edge Router CE Customer Edge
Router
6
The Software above it!
Taken from SIGCOMM 2005 Tutorial by Aman Shaikh
Albert Greenberg
7
Implications for Software Requirement

• Need to translate network requirements into
  software requirements
• Example what does five-9s of availability mean
  for software on a router?
• Need to determine performance requirement for
  tomorrow, not today
• Accommodate for growth that occurs during
  expected life-time of a software

8
Implications for Software Development

• How to write reliable and robust software?
• Systems need to work 24 x 7
• Under non-deterministic and unpredictable
  environment
• Systems should handle errors and overload
  gracefully
• Systems should support hit-less upgrade
• How to write scalable and extensible software?
• Software often lives longer than originally
  anticipated
• How to integrate software with existing systems?
• Integration with legacy software often a
  challenge
• Implication for performance
• Cater to lowest common denominator
• Exploit opportunities available for parallelism

9
Implications for Software Testing

• How to generate network-like conditions during
  testing?
• Enumerating all possible conditions is infeasible
• However, untested feature interaction can lead
  to problems
• Creating network scale is nearly impossible
• Simulators and emulators do exist, but do not
  always scale to the extent of large real networks

10
My World-View

• Knowing performance requirements upfront is vital
• But in real world, this is often not possible
• Or performance requirements change/evolve over
  time
• Due to growth
• Longer than expected shelf-life of software
• How to cope with evolving performance
  requirements?
• Analyze how software is likely to scale with
  input parameters
• Those Big-Oh and Big-Omegas are important
• Use tools to determine actual performance
• Humans not very good judge of performance and
  bottlenecks
• Have a modular design
• Cleanly separate interface and implementation
• (Re)-use existing libraries/modules
• Document, document, document,
• So that someone else can pick up the threads
• Or you yourself dont feel lost when you re-visit
  your code