CS 252 Project (Pre) Proposal

March 14, 2003

Yozo   Hida     (yozo@cs.berkeley.edu)
Paul   Huang    (pbhuang@bmrc.berkeley.edu)
Rajesh Nishtala (rajeshn@cs.berkeley.edu)

Here are three possible topics (in no particular order)
we are considering for our CS 252 project.

1. Cache design for multithreaded processors.

   Since current cache designs are mostly evaluated based on single 
   threaded code performance, there may be some design parameters
   favoring multithreaded code, such as

     + static or dynamic pooling or partitioning of cache.
     + exploiting positive interference between threads.
     + reducing negative interference between threads.
     + multithreading means cache hits may have more latency, 
       allowing more complex cache designs.
     + what to do with speculative threads.

   The plan is to investigate the performance impacts of multithreading
   on the current cache design, on a specific set of workload (such as
   sparse matrix multiply).  Then we will investigate the effect of 
   one or more changes in the cache design on the performance.

2. Effective computation over network sensors.
   
   Since each sensor comes with some intelligence, we can imagine
   a sensor network doing a distributed computation.  For example, 
   we can extend TinyDB to a simple computation to be done over network 
   (perhaps a coarse approximation of value distribution).
   We can also investigate how to efficiently gather / scatter 
   data needed by each sensor node in less-than-ideal conditions: 
   moving nodes, unreliable communication links, and faulty / malicious 
   nodes.

3. Effective computation over network attached storage

   Traditionally, many data intensive computation (such as sparse
   matrix multiply) require large I/O bandwidth just to get the data
   to perform the computation.  With network attached storage, each
   device comes with considerable intelligence, so we can investigate
   whether there are ways to efficiently process the data before 
   sending it over the network.  We will consider the effects of network
   characteristics, load balancing, storage layout, and various 
   algorithms to efficiently perform such computation.  In particular, 
   we can investigate whether this is possible under the current proposed
   iSCSI protocol and how to extend it to support such computation
   over storage.

   Alternatively: we can look at the current state of network
   attached storage and investigate its performance compared to
   things like NFS.  This will include designing a set of benchmark
   that illustrate the relative advantage (or lack of) of NAS.