1. Steven Whitham Jeremy Woods
NetSolve
Steven Whitham
Jeremy Woods

2. Architecture
A system of “loosely connected” machines, meaning they can be on a LAN or even international network
Heterogenous system, can use machines with incompatible data formats at the same time
One or many NetSolve “Agents” can exist on a NetSolve system
Each will have a view of NetSolve resources (i.e. computational servers used for calculations)
The Agent is responsible for selecting the best resource
Assuming changes to the NetSolve system are rare, eventually all agents will have the same view of the overall system
Machine application range is extending by using configuration files.
Platforms
C, FORTRAN, MATLAB, Mathematica, Java (is wonderful)

3. How It Works - Overview
NetSolve Client sends a problem to the NetSolve Agent
NetSolve Agent determines which NetSolve Resource is “best” suited for the problem and sends the problem to that resource
NetSolve System returns the result to the NetSolve Client
Communication via TCP/IP
External Data Representation (XDR) is used between hosts with incompatible data formats

4. How It Works - Details
In order to classify a describe a problem, Netsolve uses a 3-tuple of <name, inputs, outputs>
Then users must use a specific calling sequence, which is its format.
Once the client send the problem to an agent, the agent uses the network time and computation time to calculate the execution time for each machine in the system and chooses the approximate best one.
This depends on 3 types of parameters: client-dependent, static server-dependent, and dynamic server-dependent.

5. How It Works - Details Cont...
The agent ranks a list of the best machines to solve the computations, and then sends the list to the client.
The client traverses the list until the problem is solved, marking any machines that were not able to finish the request.
If the problem is not solved when the client finishes the list, the client requests a new list from the agent and this new list contains all new machines.
The result is then sent directly to the client from the machine that did the computation.

6. Execution parameters Client Dependent Static Server-Dependent
Size of data sent
Size of data received
Size of problem
Static Server-Dependent
Network characteristics between local host and machine
The complexity of the algorithm used by the machine
The performance of the machine
Dynamic Server-Dependent
Workload

7. Workload Model
The best machine is an approximation because of workload variability.
Too expensive to continuously update agent with machines workload.
Only broadcast the workload when it has changed significantly
Choice of time slice and confidence intervals extremely important.

8. Load Balancing
Best machine is determined by predicting smallest execution time T for a given problem
Time is split into time to send and receive data (Tn) and time to compute (Tc)
Tn is calculated using:
Network latency
Size of data sent
Size of data returned
Tc is calculated using
Size of the problem
Complexity of the algorithm
Performance of the server depending on current workload and its optimal performance capabilities

9. Load Balancing - Performance Model
Where: p = estimated performance P = raw performance of server w = workload n = # processors on server
Equation:
p = P x 100 x n x n + max(w x (n - 1),0)
NetSolve provides a system for creating a Problem Description File (PDF) which defines the complexity of a computational algorithm.

10. Fault Tolerance
Each NetSolve process is an independent entity (agents and machines)
The NetSolve agent automatically detects server failures if unable to establish TCP connection, and may drop the server if it does not reboot
When the NetSolve agent receives a request, it generates a weighted list of servers able to support
If the first server refuses the request, the agent sends it to each next until the request is accepted
Improper implementation of server resources will eventually cause the server to be dropped

11. Drawbacks Physical network layout Network Address Translation (NAT)
NetSolve tracks components by IP Address
Access unavailable if resource is behind NAT
Wide range of ports
Now blocked by most commercial firewalls
The core of NetSolve is still used today, but its implementation has been updated to support more applications

12. References
Henri Casanova, Jack Dongarra. Netsolve: A Network Server for Solving Computational Science Problems. April 29, 1996
Henri Casanova, Jack Dongarra. Applying NetSolve’s Network-Enabled Server, Proceedings of Heterogeneous Computing Workshop, 1998
Henri Casanova, Jack Dongarra. NetSolve: a Network-Enabled Solver; Examples and Users, Proceedings of Heterogeneous Computing Workshop, 1998
Dorian C. Arnold, Jack Dongarra. The NetSolve Environment: Progressing Towards the Seamless Grid, International Conference on Parallel Processing (ICPP- 2000), August 2000
Thomas Brady, Eugene Konstantinov, Alexey Lastovetsky. SmartNetSolve: High-Level Programming System for High Performance Grid Computing, IPDPS, IEEE Computing Society, 2006
Asim YarKhan, Jack Dongarra, Keith Seymour. GridSolve: The Evolution of A Network Enabled Solver, Proceedings of the 2006 International Federation for Information Processing (IFIP) Working Conference,

13. Thank You
Questions?