Gribble et al paper

From: Brian Milnes (brianmilnes_at_qwest.net)
Date: Wed Feb 25 2004 - 16:02:03 PST

  • Next message: Ian King: "Review: Fox et al., Cluster-Based Scalable Network Services"

    Cluster Based Scalable Network Services - Gribble et al

    The authors propose using clusters of workstations to provide incrementally
    scalable, highly available low cost internet services. They provide a layer
    of software that allows the management, transformation, aggregation, caching
    and clustering of internet services.

    Clusters are highly available and built inexpensively from commodity
    building blocks but suffer from administration problems, shared state
    problems and many services must be broken down to a smaller scale and
    reconstructed. These services can be built with less that ACID semantics
    called BASE which is basically available, soft state and eventual
    consistency. They split their clusters into management, front ends, cache
    and workers. They provide a centralized load balancer which passes out hints
    to the front end.

    They use a process peer fault tolerance in which peers restart a failed
    process. A layer is provided to handle load balancing and fault tolerance
    from worker applications. The load balancing manager is quite complicated
    using lottery scheduling and overflow management. They cached web data with
    Harvest and fixed its thundering herd startup problems.

    They then review the hotbot architecture. It's easy enough to move a
    datacenter without this level of reliabililty, I've done it three times
    live. Their self tuning algorithm is nice but a bit complicated. I simply
    stopped service above a queue length at the search engine and load balanced
    by random weights multiplied by a running average service time. Their
    dynamic addition of workstations on load spikes is very nice. When the
    Internet went down in frequently at a POP in NJ I would see huge spikes and
    had to hustle quite a few times to manually configure our spider boxes into
    service. This is certainly a much better way to build these services than I
    could provide at Lycos due to the monolithic nature of our search engine,
    but frankly it seems a bit overly complicated.


  • Next message: Ian King: "Review: Fox et al., Cluster-Based Scalable Network Services"

    This archive was generated by hypermail 2.1.6 : Wed Feb 25 2004 - 16:02:11 PST