20030410-Logistical-Ferguson

Download Report

Transcript 20030410-Logistical-Ferguson

Web100 and Logistical
Networking
Jim Ferguson
I2 Members Meeting
April 10, 2003
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Outline
•
•
•
•
What is Web100?
Web100 Diagnostic Tools
Web100 Distribution (www.web100.org)
Remember this is about Logistical
Networking
• Web100 userland Library
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
The Web100 Project
• When there is a problem, just ask TCP
• TCP has an ideal vantage point
– Can identify the bottleneck subsystem
– Already measures the network
– Can measure the application
– Can adjust itself
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Web100 Components
•
•
•
•
•
•
Kernel Instrument Set (KIS)
Diagnostic Tools
Autotuning
Widely distributed Open Source
TCP ESTATS MIB
Promote vendor adoption
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Diagnostic Tools
• User mode tools to prove core
functionality
• Template for future tool developers
• KIS validation
• Portable library to hide OS details
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
TCP Autotuning
• Supercedes today’s controls for experts
– Eliminate primary cause of the wizard gap.
– New TCP buffer management model.
– TCP just gets it right without controls.
– Paper in draft for future publication.
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Software Distribution
• Open source
– Linux kernel patches
– User mode Tools
– Contributed Software
• Active user/developer support by the
Web100 team
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
IETF Standards
•
•
•
•
TCP extended statistics MIB
Adds detailed per connection statistics
Standard TCP instruments and controls
Most recent draft submitted early March
– TSV WG work item
– Key vendors are already participating
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Kernel Instrument Set (KIS)
• TCP Instruments prototyped in Linux 2.4.
• Simple API via /proc.
• Instrument groups:
– Options and State
– IP Traffic and
Throughput
– Triage
– Congestion Events
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
– Network Path
properties
– TCP API Usage
– TCP Parameters
– WAD Parameters
Why on Logistical Networks?
• Autotuned, fast TCP for data transfers.
• Allows quick diagnosis of TCP anomolies
via access to TCP statistics.
• Autotuned, fast TCP for data transfers.
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Userland Library
• Provides useful abstraction and common
inter-application functionality
• Base abstraction is the agent (from SNMP
terminology). Other abstractions include
the group, variable, and connection.
• Includes taking snapshots and generating
log files
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Userland Library Cont’d
• Will handle different operating systems’
methods for exposing Web100 variables
(Linux’s /proc vs. BSD’s sysctl(3))
• Userland 2.0’s library’s programming
interface will change, but old applications
will not break as 1.3 and 2.0 may coexist on
the same computer
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Userland GUI
• Access, display, and control (where
applicable) values of Web100 variables
• List connections and related process info
• Triage pie chart shows the source of
congestion: Sender, Receiver, Path
• Uses GTK2/GTK1 (common to all standard
Linux distributions)
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Userland GUI Cont’d
• Display:
– continually list values of all variables
– graph those of interest
– triage analysis per connection
• Control:
– toggle auto-tuning per connection
– set tunable variables: LimCwnd, LimRwin
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications
Userland Wrap-up
• Feel free to email [email protected]
with any questions or problems with the
userland
• Feature and improvement suggestions
welcome!
National Center for Atmospheric Research
Pittsburgh Supercomputing Center
National Center for Supercomputing Applications