================================================================
Evaluation of Proposal 9905-2 PAMP
Categorized and Specialized Caching for SMP's
Hagersten
The goal of this proposal is to experiment with various shemes
to specialize data access mechanism to particular data structures in SMPs.
First data are classified according to their caching behavior and
different caching strategies are applied to different behavior.
Second, it is proposed to detect dynamically the type of behavior
for each page of data. The detection is done by software.
General comments:
1. This general approach is not new. Treating data differently
based on their caching bahavior in software was done for example
at Rice University by Bennett et al.
2. The dynamic detection part is, based on my experience, a very hard
problem. Many times, a dynamic algorithm works for some program but
fails for others. Whereas the approach is intellectually appealing,
computer manufacturers rarely rely on unpredictable mechanisms for
enhancing performance. Furthermore the success of the dynamic
prediction scheme is predicated on the success of the first part,
i.e. that classifying data is actually beneficial.
3. It is not clear that by improving the caching behavior of data
real-time applications will necessarily benefit.
Relevance to PAMP:
The goal of PAMP is to apply SMP technology to
real-time applications. Therefore improving the performance and
the PERFORMANCE PREDICTABILITY of SMPs falls within the perview
of PAMP. Classifying data is a possible way to achieve that goal.
Synergy:
Hagersten has academic and industrial contacts both
in Sweden and in the US. I am not sure that these contacts
will be eager to work on this project.
Industrial relevance is average. The work proposed here may possibly
have an industrial impact.
Academic qualities:
The basic idea is not new but is so appealing
that it probably deserves another try. Hagersten has the qualities
to make this happen.
Misc.:
The letter from Ericsson is VERY weak. Basically
Ericsson just want to "stay in touch".
Recommendation:
I would fund the first part (one Phd student) to classify data and see the
impact on real-time processing of real-time applications. Then I would
consider funding the second part.
================================================================
Evaluation of Proposal 9905-4 PAMP
Software Distributed Shared Memory-- New Applications and Scalability
Brorsson
This is a very ambitious proposal whose goal is to make software DSMs ("S-DSMs")
work in large-scale configurations (256 processors?) and for real-time applications.
There is no indication
that this is possible and Brorrson does not seem to have any idea as
to how he is going to make this happen. All evidence points to the fact
that speedup saturates after 8-16 processors for S-DSMs in the context of
scientific applications. No argument is made in the proposal that this situation
could improve for real-time applications.
Additionally, the proposal lacks focus, has an ill-defined methdodology, and
does not present a compelling case that success is possible.
1. It lacks focus. I am not sure how success can be measured on this proposal.
So many things are promised that Brorrson cannot possibly succeed on all of them.
Actually what is promised is not that clear, but I gathered the following
throughout the proposal (correct me if I am wrong)
* Database applications and websevers will be studied in the context
of S-DSMs (on actual prototypes of the hardware, database engine and applications)
* A 4-processor hardware prototype will be built in the 1st year; then 1 16-processor
prototype in the second year; then a 256-processor prototype thereafter.
*A large-scale surveillance system will be used as a benchmark application.
Each of these three parts are MAJOR endeavors.
2. The methodology is ill-defined. Basically the methodology relies on
"heavy" prototyping, which is the most time consuming and (by far) expensive
way to get to any result. Even if the prototyping is successful, the prototypes
cannot probably yield the kind
of insight that is needed in this proposal. The use of one FPGA in the prototype
is mentioned. How it can be used is unclear. Why do we need to build a hardware
prototype for what is essentially a software project????
3. No sign of success. Not one sentence in the proposal argues that S-DSMs could do
better on real-time applications than on scientific applications, except that "we
have to look at them". No new architecture concept is proposed to attempt to
dispell the poor speeup of S-DSMs.
Relevance to PAMP:
The promise to look at databases, web servers and surveillance
applications makes the proposal somewhat relevant to PAMP.
Synergy:
Not much evidence of synergy in the proposal
Industrial relevance is average. The work proposed here may possibly
have an industrial impact.
Academic qualities:
As an academic exercise the goal of this proposal (i.e. scalable
S-DSMs) is worthwhile pursuing. Howevr the proposed methodology is not cost-effective
in terms of research. Its chance of sucess is very slim.
Misc.:
The "letter" from Axis communication commits to pay for the hardware for a
260 processor system. This is a very significant commitment.
However the copy of the letter that I have is not signed and does not
have any official logo.
Recommendation:
Focus and reformulate.