Sprungmarken

Servicenavigation

Hauptnavigation

Sie sind hier:

Hauptinhalt

A Spinning Join That Does Not Get Dizzy — PVLDB Reviews

Reviews for paper A Spinning Join That Does Not Get Dizzy, submitted to PVLDB (Proc. of the VLDB Endowment).

Overall rating: reject

Reviewer 1

Overall Rating

Weak Accept

Novelty

Some New Ideas

Technical Depth

Competent, but a little more would be good

Motivation and Impact

Well motivated

Presentation

Very well organized and written

What is the key contribution of this paper? Please write a brief justification of your overall recommendation

This paper proposed a ring-structured distributed DBMS that uses remote direct memory access (RDMA) to transfer data among nodes. It also proposes a distributed join technique for such a DBMS.

What are some strong points of this paper? Please number them S1, S2, etc.

S1: The paper is well-organized and clearly written.
S2: I liked that the paper touches on some of the system issues (e.g., the use of pre-allocated buffers) involved in making efficient use of RDMA.

What are some weak points of this paper? Please number them W1, W2, etc.

W1: The contribution is small. (See C1)
W2: The paper needs a better discussion of previous ring-structured systems (especially DataCycle, also broadcast disks), so that it is clear what is new here. What are the significant differences between that work and this?

Detailed Comments

C1: This paper is pretty thin. The join algorithms themselves are derived in a very straightforward way from MonetDB. The experiments are well-thought out but limited in scope, and the results are not surprising. I would have liked to see this paper cover more ground, either algorithmically or experimentally.

C2: Figure 9 was not clear. What are the local (black) bars with which cyclojoin is being compared.

Are there grounds for summary rejection (technical error, plagiarism, out of VLDB scope, ...)? Please check this box and provide confidential comments.

No

Reviewer 2

Overall Rating

Weak Reject

Novelty

Nothing New

Technical Depth

Enough Technical Material for a VLDB Paper

Motivation and Impact

Important and/or Impactful

Presentation

Adequately organized and written

What is the key contribution of this paper? Please write a brief justification of your overall recommendation

The paper revisits parallel join algorithms in light of recent progress in fast networks using Remote Direct Memory Access (RDMA). Since RDMA yields very fast transfer rates with low CPU overhead, moving data to be joined over the network becomes very fast and can be exploited to completely eliminate disk access, using multiple machines’ main memories. After making that point, based on a ring network of multiple machines, the authors propose a cyclo-join algorithm that exploits the cycling flow of one operand relation over all machines, each holding a relation’s piece, to distribute join processing. The experimental study shows that, with such approach, the join of large relations can be processed as multiple main-memory joins and thus very fast.
The key idea has already been published by the authors in a recent paper.

What are some strong points of this paper? Please number them S1, S2, etc.

S1. The paper is well written and easy to read.
S2. The analysis of the potential of RDMA for database processing, and the new trade-offs between CPU, disk and network performance is thorough and well done.
S3. The proposed architecture (ring network) and distributed join algorithm (block nested loop) are very simple and easy to implement and deploy.

What are some weak points of this paper? Please number them W1, W2, etc.

W1. Most of the content of the paper, in particular, the main idea (including ring network and cyclo-join algorithm) have been published at the 2009 DaMon workshop.
W2. The ring network topology which is the basis for the contribution has well-known fault-tolerance problems, which does not make it a good fit for the cloud computing environments which is targeted.
W3. The distributed join algorithm is real simple, block nested loop, and it is not clear whether other hash-based parallel join algorithms would be better.

Detailed Comments

Most of the content of the paper, in particular, the main idea (including ring network and cyclo-join algorithm) have been published in the DaMon workshop, whose proceedings are published by ACM. Besides some optimizations, there is little added value in the submitted paper.

Ring network: this is a very simple topology, with well-known fault-tolerance problems (e.g. one machine down along the ring creates problems for the entire network). Since one motivation of the paper is cloud computing (typically with very large number of machines), this can be a major issue. Other network topologies such as bus or tree might more attractive in this context.

Distributed join algorithm: the main goal is to reduce response time for ad-hoc queries, hence the minimal requirements on data placement over the multiple machines. However, each machine still needs to join the local fragment with the entire relation. The authors show that the local join can be done as fast as possible (using radix or sort merge join) in main memory, possibly with multiple cores. However, there may be other combinations of network topology / distributed algorithm (e.g. tree-based as Teradata’s bynet with hashing) which could avoid having to perform a local join with an entire relation.

Main memory utilization in multiuser environment: in a cluster system where tasks are allocated to any machine in the network for load balancing, it seems that the proposed join technique which has strong main memory requirements (to make sure there is no disk access during local joins) can affect the multiuser level, or lead to users starvation for main memory.

Are there grounds for summary rejection (technical error, plagiarism, out of VLDB scope, ...)? Please check this box and provide confidential comments.

No

Reviewer 3

Overall Rating

Weak Accept

Novelty

Creative

Technical Depth

Enough Technical Material for a VLDB Paper

Motivation and Impact

Very Important/Impactful

Presentation

Very well organized and written

What is the key contribution of this paper? Please write a brief justification of your overall recommendation

The paper proposes a distributed join algorithm that uses a ring of join nodes and pushes one of the two joining relations through the ring, while the other relation is distributed across all nodes. A particular novelty is that the communication is done via remote direct memory access instead of traditional point-to-point communication such as TCP/IP.

What are some strong points of this paper? Please number them S1, S2, etc.

S1 Exploitation of a very interesting new hardware support that makes communication much faster than traditionally assumed

S2 relatively simple setup so that it might actually be used; also practical for cloud computing environments.

S3 implementation and evaluation on a real system with a hardware configuration reflecting current setups within data centers

What are some weak points of this paper? Please number them W1, W2, etc.

W1: As the paper mentions there have been many distributed joins proposed before. Instead of comparing with a single node, it would have been more interesting to compare the ring approach with more traditional distributed joins in regard to two aspects. What is the fundamental difference of the ring approach compared to existing distributed join algorithms? A short description of the best known existing distributed join algorithms would help in this context. It would have also be interesting to compare the ring approach experimentally with one of the existing approaches -- adjusted to also use RDMA.

My problem here is probably that I don't quite see why the ring approach is particularly useful with fast communication and why the existing distributed algorithms would not benefit from faster communication.

W2: The analysis does not seem to consider the cost for reading the data from disk into main memory and the initial overhead to decide on distribution of the inner relation. Depending on the disk setup, loading the data from disk into either a single main memory or into the distributed setup might be comparable or provide a further speed-up for the distributed architecture. That would have been a further interesting point to analyze.

W3: The comparison of RDMA compared to TCP/IP should definitely go into the main part of the paper. At the end, this paper is motivated by using faster communication infrastructure instead of traditional TCP/IP. I was a bit disappointed to only see a speedup of at most 2.

Detailed Comments

The paper uses very specialized joins that also take advantage of the multi-core architecture. A small discussion of what could be expected if standard vanilla joins, as implemented in current DBS, would be used.

Suggested Revisions. Are there changes you think would make this a better VLDB paper? Please number your suggestions R1, R2, etc.

They refer to the points mentioned in section "Weak Points".
Although some of the requested further evaluations will probably incur a high overhead.

Are there grounds for summary rejection (technical error, plagiarism, out of VLDB scope, ...)? Please check this box and provide confidential comments.

No

Related Information



Nebeninhalt

Kontakt

Prof. Dr. Jens Teubner
Tel.: 0231 755-6481