Suman Nath
----------

Dynamic Function Placement for Data-intensive Cluster Computing (ABACUS)
=================================================
Optimally partitioning of application and file-system functionality within 
a cluster of clients is very difficult because of dynamic variations in 
application behavior, resource availability and workload mixes. Many of 
these characteristics cannot be predicted before running the application, 
and some of them (like availability of a server) may change dynamically 
during execution. Hence universal solutions do not exist and any attempt 
for such a solution will lead to sub-optimal and often catastrophic 
performance.

The ABACUS framework strives to achieve an optimal solution for a 
restricted class of applications viz. data-intensive applications. It uses 
a dynamic approach where the function placement is continuously adapted by 
run-time system based on resource usage and availability. It demonstrates 
that placement can be decided based on black-box monitoring of application 
objects, in which the system is oblivious to the function being implemented.

Abacus run-time system consists of a migration and location-transparent 
invocation component, and a resource monitoring and management component. 
The first component is responsible for the creation of location transparent 
references of mobile objects, for the redirection of method invocations in 
the face of object migrations, and for enacting object migrations. The 
second component collects statistics and decides on which object to place 
where.

A prototype of Abacus has been implemented and experiments indicate it 
achieves 2-10X performance improvement and successfully adapts to both 
long-term and short-term runtime variations.

However, the success of Abacus relies critically on the effectiveness of 
the analytical model used for taking decision on placing the mobile 
objects. It is unclear how robust the model is, how quickly it adapts to 
variations, and whether the model itself is static or can be modified 
dynamically. The benchmarks used to evaluate Abacus seem to be synthetic.

The message I take from the paper is that where there is no 
"one-size-fits-all" solution, we should avoid unnecessary and expensive 
generalization. (Do not Generalize!!)