2005 © switch pert – beyond fat pipes simon leinen

16
2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen <[email protected]>

Upload: sarina-burkes

Post on 01-Apr-2015

227 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH

PERT – Beyond Fat Pipes

Simon Leinen <[email protected]>

Page 2: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 2

Contents

• End-to-end performance initiatives and PERT history• Example PERT activities: GN2-PACE and SWITCH• Lessons learned so far• How Grid developers and users can benefit from PERT

Page 3: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 3

End-to-End Performance Initiatives

Motivation• Observation: Backbone improvements not reflected at end users• Frustrations with differentiated-QoS mechanisms

Ideas for improvement• Education (close the “wizard gap”)• Measurement infrastructures

– In the network, e.g. GN2-JRA1– In the hosts, e.g. Web100

• Human support infrastructure → PERT

Page 4: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 4

Example PERT Initiatives: GN2 PERT

GN2 “PACE” (formerly known as “SA3”) PERT• Oriented towards European projects—including Grids• Large virtual team with participants from many NRENs• Weekly “Case Manager” duty, rotating between organisations• http://www.pert.geant2.net/ (still being built)• http://kb.pert.switch.ch/ (Pilot Knowledge Base)

Page 5: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 5

Example PERT Initiatives: SWITCH PERT

• Includes support for performance of the “commodity” Internet• Small team from our NOC/network engineering group

Other NRENs are building PERT groups as wellbut these are the two I’m most familiar with…

Page 6: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 6

DEISAThroughputReduction

Example GN2 PERT Issue• Iperf (single-stream TCP) measurements between the German and

French DEISA sites usually achieve 900 Mb/s.• DEISA traffic uses Premium IP inside the GEANT backbone.• In parallel, there are 2-3 Gb/s throughput tests between Karlsruhe

and CERN, using LBE (Less Than Best Effort) inside GEANT.• Although there shouldn’t be an impact, DEISA measurements drop

to 400 Mb/s during the Karlsruhe/CERN tests.

Page 7: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 7

DEISAThroughputReduction Issue (2)

DEISA Iperf test traffic

Normal DFN-GEANT traffic

Page 8: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 8

DEISAThroughputReduction Issue

Possible Avenues Towards Resolution:• Missing QoS configuration on DFN's router to GEANT• Try TCP variants that fill the pipe more easily (e.g. BIC, FAST)• More measurements (delay/loss/reordering when with bg. traffic?)• Is this a practical problem at all?

Page 9: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 9

Slow access to download.microsoft.com

Example SWITCH PERT Issue• Users at site X observe low download speeds from software

distribution site (30-70 kbps)

Page 10: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 10

Download Issue (2)

• First tests from a test host close to X result in good rates (30-40 Mbps) → probably an issue within site X, right?

$ wget http://download.microsoft.com/.../dotnetfx.exe

Resolving download.microsoft.com... 195.176.255.136, 195.176.255.135

Connecting to download.microsoft.com[195.176.255.136]:80... connected.

HTTP request sent, awaiting response... 200 OK

Length: 24,265,736 [application/octet-stream]

100%[====================================>] 24,265,736 4.49M/s ETA 00:00

13:44:22 (3.54 MB/s) - `dotnetfx.exe' saved [24265736/24265736]

Page 11: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 11

Download Issue (3)

• However, subsequent tests reproduce rates similar to what the customer observed.

$ wget http://download.microsoft.com/.../dotnetfx.exe

Resolving download.microsoft.com... 212.162.0.30

Connecting to download.microsoft.com[212.162.0.30]:80... connected.

HTTP request sent, awaiting response... 200 OK

Length: 24,265,736 [application/x-msdownload]

12% [===> ] 2,968,483 7.37K/s ETA 51:32

• Further investigation shows that mapping of download.microsoft.com to actual hosts is highly variable, and the performance of the servers mapped to is even more variable.

Page 12: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 12

Download Issue (4)

• Because SWITCH doesn’t entertain business relationships with Microsoft of its content distributors, suggested that our customer complain to Microsoft (of which they are also a customer).

• Although technically a PERT failure, customer seems happy.

Page 13: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 13

Lessons Learned So Far

• It’s hard to establish contacts with the customer• It’s easy to “lose” the customer• Performance expectations are often unclear• Very hard to close an issue

– Lack of criteria for problem resolution

• Should think in terms of bottlenecks rather than “problem locations”

Page 14: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 14

How Grid Community Can Benefit Today

PERT Knowledgebase (http://kb.pert.switch.ch/)• Lots of performance-related information• In the process of being builtContact the PERT in case of performance problems• If it happens when using GEANT/GN2• After other avenues have been explored (campus system and

networking support)• Through your NREN or (for groups such as DEISA) directlyNew: PERT Consultancy Service• Ask PERT for help with future applications/tests (proactive)

Page 15: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 15

How the Grid Community can Help PERT

• Follow the project and provide constructive criticism• Volunteer as GN2 PERT Subject Matter Expert (SME)• Make the concept known within your communities• Bring interesting cases before the PERT

Page 16: 2005 © SWITCH PERT – Beyond Fat Pipes Simon Leinen

2005 © SWITCH 16