1. Muhammad Murtaza Yousaf, Michael Welzl
A Reliable Network Measurement and Prediction Architecture for Grid Scheduling
Muhammad Murtaza Yousaf, Michael Welzl
NSG Team, Distributed and Parallel Systems Group,
Institute of Informatics, University of Innsbruck, Austria.

2. Motivation High reliability = main goal for these predictions!
Selection of optimum site is crucial for Grid Scheduler.
“A scheduler always needs to know how long file transfers will take.”
High reliability = main goal for these predictions!
Traffic changes a lot; routing doesn’t
thus, measure / predict path (not traffic) properties!
=> represent minimum transfer delay (ideal case)
Grid Scheduler
Grid
Site – 1
Site – 2
Grid Site
Source

3. Design Consideration
We are looking for a measurement and prediction system which provides:
Reliable estimates, with
Small overhead, and it must be
Scalable, with
High availability.

4. Architecture Description
Five basic components:
Client
Monitor
Active Prober
Information Base
Information Manager
Monitor
Information
Manager
Client
Active
Prober
Base
High level architecture of the system

5. Architecture Description (Cont…)
Information
Manager
Network
Interface
Sniffer
Extractor
Filter
Internal architecture of the Monitor
Monitor has sub components:
Sniffer
Monitors and captures traffic from a network interface.
Filter
Packets captured by the Sniffer may belong to many flows
separates packets which belong to a particular stream.
Information Extractor
Receives a filtered stream
generates reliable estimates by using our measurement methodology.

6. Architecture Description (Cont…)
The Active Prober
injects some traffic for the measurement purpose,
generates estimates as required.
The Information Base is to store measurements.
The Information Manager
stores and retrieves information from Information Base with a timestamp.
Interacts with the client.
Invokes Active Prober, if
Information is too old
Network path is lightly loaded
Information is missing
Client behavior requires more reliable information.

7. Detection of Shared Paths
Detection of shared paths and shared bottlenecks will enhance the usefulness of our predictions A B C α β E D

8. Bandwidth Estimation Technique: Packet Pair
Two packets sent back-to-back
Looks Active, but possible in Passive
Interspacing after bottleneck
Inverse of bottleneck bandwidth
Preserved at non-bottleneck routers.
Pi-1
Direction of flow Pi ∆ = s b
Packet Pair (Pi-1 and Pi), before and after bottleneck

9. Conclusion and Future Work
A measurement system to estimate network path characteristics in a grid.
Passive (mostly).
Packet Pair based Measurement Technique (relationship of ACKs and corresponding data packets).
RTT Estimation
Currently we are working on a preliminary version of the Information Extractor component of the Monitor for a reliable estimate of the path capacity (real-life and simulation).
After an accurate and reliable estimate of bottleneck bandwidth, our target is the detection of shared bottlenecks because this knowledge can enhance the usefulness of our predictions.
For experimental results we plan to use our system in Planet Lab, and our ultimate target is to deploy it on Austrian Grid.