================================================================
Evaluation of Proposal 9905-2 PAMP
Categorized and Specialized Caching for SMP's
Hagersten
This proposal aims to incorporate selective caching techniques into a
commodity SMP architecture, with the goal of using commodity SMPs
instead of customized processors in environments with large data sets
and real-time needs like telephone switches. The project will be
performed in close collaboration with architects at Ericsson, who have
expertise in the design of customized processors for this area.
This is a promising proposal. The overall idea---namely, making
(hopefully) small adjustments to commodity SMPs so that they can be
used effectively in such real-time applications---is a good one to
pursue since it may lead to the leveraging of the commodity
performance curve and investment in this important application
area. The team seems very appropriate: the Ericsson collaborators
have expertise in custom-processor design and in the applications, as
well as access to appropriate workloads to drive the design and
evaluation, and Prof. Hagersten is an internationally-recognized
expert in the design of commodity-based architectures and systems with
substantial industrial experience in both this area as well as
real-time systems (at Ericsson). In addition to designing successful
systems in industry, he has made many valuable contributions to
computer architecture research as well.
The weakness of the proposal is in its lack of specificity.
Customized caching techniques have been proposed for general-purpose
computing and for special purpose computing for quite some time, as
the proposal recognizes. While it is natural to see it being effective
for special-purpose systems (e.g. media and streaming applications),
the case for general-purpose systems has not appeared to justify the
hardware complexity and intervention in commodity systems. The key
strength of this proposal is its aim to target changes to commodity
systems to make them more appropriate for this environment. However, the
challenge is that these are general-purpose systems that are used in
many more environments than real-time ones, so the changes must indeed be
small enough to be acceptable and important enough in terms of
performance improvement to influence the commodity industry. The
proposal does not point toward any specific novel techniques toward
this goal. While this is presumably the task of the research, I like
to see some key ideas or directions in a proposal.
I recommend funding the proposal based on the strengths and
experience of the personnel and the overall goal it aims to pursue.
================================================================
Evaluation of Proposal 9905-4 PAMP
Software Distributed Shared Memory-- New Applications and Scalability
Brorsson
This proposal aims to study three major areas: (i) the design of
software distributed shared memory (S-DSM) systems for emerging web
server and database applications, (ii) the design of S-DSM for
embedded systems in surveillance applications, and (iii) the scaling
of these systems.
The first area is an increasingly relevant area of study. However,
the authors do not make a strong case for the importance of S-DSM in
such systems. Key aspects of web servers, perhaps more important than
performance optimizations, are availability and dealing with
faults. It is arguable that these are easier to achieve with message
passing than with a shared address space, and it is not clear so far
that web servers are complex enough in their treatment of data to
argue strongly for a shared address space. This lack of motivation
can be attributed to length restrictions in the proposal, but it would
have been good to have seen a stronger case made.
The third area, scalability of S-DSM systems, is important. The
authors have been doing research in S-DSM systems, so they should be
qualified to do this research.
The second area, S-DSM systems for embedded systems, is one that I did
not understand fully from the proposal. It was not clear what new
issues arise for S-DSM from the fact that embedded systems are being
used, or even whether the S-DSM is being used across the "sensor"
nodes or only on the database server (I assume the former). Is it
simply that surveillance is a new application that is appropriate for
S-DSM, or that somehow the embedded nature of the processors raises
new issues. Studying new applications for systems is good. However, if
the case is being made for new research issues, they are not
articulated.
Overall, this proposal seeks to do research in generally important
areas, so good things may emerge. Its negative is that it does not
make its case very strongly or present many new ideas.
================================================================
ARTES: 9905-7
TITLE: Methods for Integration of Heterogeneous Real-Time
Services into High-Performance Networks
AUTHORS: Magnus Jonsson and Bertil Svensson
This proposal seeks to develop methods to use an existing special
control-fibre in a fiber-optic ring interconnect for a multiprocessor
to provide soft/hard real-time and quality of service guarantees to
applications.
The goal is fine, and such research is interesting to perform.
However, the current proposal has some weaknesses. First, and most
important, no ideas are stated for how the control-fibre will be used
to provide the higher-level or more diverse QoS features the proposal
seeks to provide. That the control fibre can be used for such
purposes is presumably the purpose of the control fibre in the first
place, which is previous work. Nor are any concrete instances
provided of the needs of applications that might be satisfied in this
way. Thus, there is no reason provided for a reviewer to feel that the
investigators will have a high likelihood of success. Second, the
case for optical networks as the networks of choice for
*multiprocessors* with real-time constraints is not made well. Yes,
they have additional bandwidth, but they have not yet proved ready for
multiprocessors in general due to the latency/overhead of the
conversion between optics and electronics. This is okay, since the
research is likely to be valuable anyway, but the investigators should
have addressed this. Third, providing real-time guarantees in the
interconnect fabric is not nearly enough to provide end-to-end
real-time guarantees to an application. This is especially true in
multiprocessors where the dominant cost of communication, as well as
the dominant contention for resources that can make timing
unpredictable, is not in the network fabric but at the end-point nodes
of the communication. The authors should have paid some attention to
these issues, especially the first, even in a short proposal.
The specific area of the proposal seems relatively untrodden and good
methods may come of the research, but the case is not well made.
================================================================