Sei sulla pagina 1di 1

In this paper, the authors present the viability of machine learning in the context of branch prediction.

Two
tipes of static branch prediction and one type of dynamic branch prediction solutions are explored.
The first solution, static branch prediction with neural networks, is presented and its misprediction rate is
reported. The disadvantages of this solution consist in the high number of elements needed for the training
set.
The second solution consists in static branch prediction with genetic algorithms. A historic introduction of
the technique is followed by a succint explanation of its guidelines. A table is then shown, that compares the
performances of this solution with traditional one-bit and two-bit predictors. The authors follow by listing
the disadvantages: the difficulty to implement due to very deep tree structures and the impossibility for the
predictor to evolve further after manifacture.
The third sections deals with dynamic branch prediction with perceptrons. The history of perceptron research
is briefly summed up and followed by a technical explanation of the implementation, advantages and
disadvantages of the solution compared to Gshare and B-Mode predictors. The autors, then, present the
improvements on perceptron branch prediction proposed by D. Jimenez.
In the conclusion, the authors cite D. Jimenez in saying that the amount of area needed for the
implementation of perceptron branch prediction is not justified by a sufficient increase in performance.

One of the strong points of the paper is the relevancy of the topic. Branch prediction is an important
technology and machine learning is a rapidly evolving field. This results in a very interesting paper.
Furthermore, the summary of the topics, when given, is solid and easy to understand.

Moving on the weak points:


1- the paper doesn't follow the guidelines, since it doesn't have a comparison section. There are some
comparison arguments in the summary sections, but they can all be found in the original papers and only
compare the solutions with more traditional branch prediction solutions. This is to say that no actual effort
was done to compare the performances, area or any other parameters of the machine learning solutions
between each other;
2- in the abstract, the authors claim that: “The dynamic variant, the perceptron, outperforms all other
methods”. This claim, however, was not explored in the paper in a proper section and is left unsubstantiated;
3- the conclusion only considers the perceptron predictor and not a word is given on the other two solutions
and on their viability;
4- the paper doesn't include an explanation for what a perceptron is, and even encourages the reader to 'self-
study' based on the original paper material. I understand that this criticism may not be shared by everyone,
because, as the authors say, the perceptron itself is not the main topic of this paper. In my opinion, however,
this is particularly faulty, because with a missing comparison, most of the evaluation should be based on the
summary work. This summary was, instead, avoided, even though I reckon that it was possible to do in a few
paragraphs at most and certainly there was enough space to do so within the page limit;
5- the authors, use a large number of papers for citation (12), without clearly stating the papers actually
selected and approved for the comparison work (that should be 5 maximum). This made it difficoult for me
to judge the validity of the papers because I could not discern those from the mass.

I consider all of the problems above to be fundamental, and reduce the strenght of the paper to just summary
and citations of other works. There are, however, some minor issues. The last sentence of the introduction
was left open, and contains a hint for a performance oriented comparison not delivered. Moreover, there are
some typos in the document, but in low number and of low importance.

In my opinion, this paper deserves a reject, because it fails to follow the guidlines in multiple ways. It gives
unsubstantiated claims and misses the comparison section completely, while the technical presentation
misses sections that could be explained further.

Potrebbero piacerti anche