User Tools

Site Tools


en:problems:dweb:tuning
 
 

Autotuning

The performance of distributed systems is partially dependent on system configuration parameters like session time (how long the conversational state with one node is maintained across multiple requests), maximum number of connected nodes (size of neighbourhoods), cache pool size (for storing temporary variables) and data store size.

  • The values of configuration parameters are usually kept in plain text files or as XML text.
  • Parameter tuning requires little extra cost compared with demanding hardware changes.
  • Parameter values significantly impact system performance.

Configuration problems

  • If the session time is set too short (reducing the potential for malicious attacks), nodes have to more frequently connect, leading to connection overhead.
  • Increasing the number of maximum nodes to connect to and cache pool size increases system throughput, but cannot guarantee satisfactory response time because nodes have to deal with more data, maintenance states and temporary information.
  • In p2p systems a major difficulty in optimising performance is that decentralised data and distributed control resides on different nodes, requiring dynamic distributed system optimisation. And as the internet evolves and new devices are added, so do load balancing requirements. Dynamic and random events in complex systems are difficult to predict using precise mathematical models.
  • Determination of the proper configuration parameters is quite time consuming using a black box search model for optimisation.

Heuristics for autotuning

Heuristics on a node's neighbourhood can possibly be used for autotuning. A heuristic function that will rank all the possible alternatives for a better tuned configuration based on the available information. It helps the algorithm to select the best route out of possible routes to find a better configuration (if exists). Each agent then has its own agent attributes to propagate to its neighbours.


en/problems/dweb/tuning.txt · Last modified: 2020/07/03 19:25 by Digital Dot