ns2 project in goa
ns2 project in goa recall that the match effort for base rules is distributed by
heuristically choosing a restriction attribute (RA) for
each rule. A restriction predicate on the RA restricts
the match effort per rule at each base rule processing
site, thereby distributing the work done in base rule
matching. To achieve this, the metarules are analyzed
at compile time to determine the restrictions on base
rules as follows. Base rule RAs are ns2 project in goa chosen so that at each site, only
those instances are generated that are relevant to
each other with respect to the set of metarules. However, ns2 project in goa at the same time, we must ensure that
all possible instances are generated over all the
sites. The abovens2 project in goa two goals may be difficult to satisfy
simultaneously. Generating relevant instances according
to the metarules may require ns2 project in goa generating more instances than in the ordinary case where we
use restriction predicates that divide the range of
each base rule RA according to the processing potential
of the sites, to get match time reductions
per rule and uniform completion times over all
sites. This is so, since the distribution of relevant
instances may not agree ns2 project in goa with that determined by
the algorithm that is only concerned with dividing
up the match effort over all sites. The PARADISER system is now operational with
base rule processing being fully distributed and
metarule processing being carried out at a single MRP.
Initial experiments have shown that for large data
sets, the BRPs perform well under dynamic load balancing
to compute base rule matches in parallel and reduces the match time thus allowing the system to
scale. However, the single site ns2 project in goa MRP emerges as a bot,-
We have begun work on distributing the metarule
processing as well. ns2 project in goa Our current effort is to provide
distributed metarule processing using the full distribution
(FDM) scheme. Our initial experiments point
to typical patterns in the behavior of a rule-based system
as execution progresses. One such pattern is that
after several cycles, only a few metarules are typically
active, since many of the base rules relevant to the
bulk of the metarules do not produce instances after
the first few cycles. Under the FDM scheme, each
metarule is assigned to a distinct processor. Thus, in
a realistic situation, many of the ns2 project in goa MRPs may be inactive.
We have developed protocols that detect whether
or not a particular MRP will be active at a given cycle.
It is then possible to distribute the processing
load of the active MRPs over all available MRPs, i.e.,
we use the resources of the ns2 project in goa inactive MRPs whenever
possible. The load distribution is based on the assumption
that at each cycle, the entire conflict set is available at every MRP, since they can all collect all
of the instances as they are broadc.ast by the BRPs.
Load distribution then reduces to following a protocol
that determines how active ns2 project in goa MRPs will claim the
resources of inactive ones, and then executing some
variant of a popular parallel join algorithm adapted
to our particular needs. In future work, we plan to conduct comparative
studies of the various approaches to distributed control
as detailed in this paper with actual implementations within the PARADISER architecture. We also
intend to explore alternative and useful control specifications of practical importance. In [9], a number
of PARULEL programs are studied and some prelirninary
ideas are presented on how ns2 project in goa to improve the expressivity
of the metarule construct of PARULEL. These
are important open problems that will be the focus of
future work.