Reviews for paper Frequent Item Computation on a Chip, submitted to TKDE (Special Issue on Best Papers of ICDE 2010).
Overall rating: accept with no further changes
Accept With No Changes
The new version of the paper clearly meets the "new material" criteria.
The paper is mostly very well written (just a handful of awkward sentences here and there), and the experiments are mostly quite convincing and explained in enough detail to allow replication.
I do think that we will see a lot more papers using FPGAs or GPUs to speed up data mining algorithms in the coming years, and I expect this paper to be well cited.
Minor comments.
"Abstract-Computing frequent items is an important problem by itself and as a first step for several data mining algorithms" It is important as a sub-routine in several data mining algorithms, but not necessarily the first step.
You have 12 occurrences of "In this paper.", but were else would it be?
"In addition, and not least important these days,.." Awkward.
"The idea of systolic arrays was also picked up by Baker and Prasanna"
"The idea of systolic arrays was also considered by Baker and Prasanna"
Research/Technology
Very Relevant
I do think that we will see a lot more papers using FPGAs or GPUs to speed up data mining algorithms in the coming years, and I expect this paper to be well cited.
Appears to be - but didn't check completely
Yes
References are sufficient and appropriate
Yes
Satisfactory
Easy to read
Excellent
Accept With No Changes
As the original reviewer of the ICDE'2010 version of this paper, I am glad to see that the authors had continue to perform very detailed studies of the problem and address some of the comments that are made in the ICDE'2010 reviews. I therefore proposed to accept this paper without future changes.
Research/Technology
Very Relevant
This paper is an extension of an excellent ICDE'2010 paper discuss how frequent items can be computed on FPGA. This paper extended the original version in various ways. First, a new section on answering queries is included to explain how queries can be answered efficiently during the FPGA computation. Second, it discuss about energy efficiency and show that their implementation can process high data volmes directly off a giagabit Ehternet connection which consuming a fraction of the energy resources. Finally, by using a new version of their code base, this new version provide performance that improve over the conference by 30 percents and studies can be performed over a large range of problem sizes.
Yes
Yes
References are sufficient and appropriate
Yes
Satisfactory
Readable - but requires some effort to understand
Excellent
Accept With No Changes
While the problem of finding frequent items is well studied for software solutions on general purpose CPUs, exploiting FPGAs for this purpose has not been studied before. In fact, research into efficiently using modern FPGA technology for datamanagement purposes is still in its infancy. This work provides a significant contribution. The proposes FPGA solutions provide significant performance improvements over software solutions in general purpose CPUs, while consuming significantly less energy. In addition, the paper provides a concise description of FPGA technology and a very educative introduction how "thinging the FPGA way" is required to design highly efficient FPGA solutions to data management problems.
Research/Technology
Very Relevant
This very well written paper provides an educative show case of how to exploit FPGA technology to provide new high-performance solutions to data management problems. Using the problem of finding frequent items in large data stream as example, the author present three alternative FPGA solutions. Starting with a straight-forward FPGA-adaptation of a classical software solution for general purpose CPUs, the author gradually develop improved solutions that exploit FPGA characteristics.
While the problem of finding frequent items is well studied for software solutions on general purpose CPUs, exploiting FPGAs for this purpose has not been studied before. In fact, research into efficiently using modern FPGA technology for datamanagement purposes is still in its infancy. This work provides a significant contribution. The proposes FPGA solutions provide significant performance improvements over software solutions in general purpose CPUs, while consuming significantly less energy. In addition, the paper provides a concise description of FPGA technology and a very educative introduction how "thinging the FPGA way" is required to design highly efficient FPGA solutions to data management problems.
Yes
Yes
References are sufficient and appropriate
Yes
Satisfactory
Easy to read
Excellent