Poster Title:
Towards Service Pool Based Approach for Services Discovery and Subscription
There are many function identical web services in the internet or some large-scale organizations. They provide consumers with more choices according to their personalized QoS requirements. However, in current web service discovery and subscription, consumers pay too much time on manually selection and cannot easily benefit from the wide QoS spectrum brought by the proliferating services. In this paper, we propose a QoS-aware discovery and subscription approach to free consumers from time-consuming human computer interactions while helping them negotiate QoS with multiple service providers. The core of this approach is the service pool, which is a virtual service grouping function identical services together and dispatching consumer requests to the proper service in terms of QoS requirements. Based on our previous work on the service pool, this paper does two main contributions: one is we investigate the pool construction mechanisms by the similarity retrieval and formalize the representation of the service pool; another is we propose a formalized QoS model from consumer's perspective, and design a QoS-aware discovery algorithm to select the most adequate provider from the service pool according to consumer's QoS requirements. ___________________________________________________________ Authors Declaration

This paper is originally submitted as a regular research paper to WWW2007. It was reviewed by three reviewers(two reviewers accept, and one rejects). We carefully cut the paper to current version according to the reviewer's advice. The following is the reviewer's advice ________________________________________________________

-------------------- review 1 --------------------


----------------------- REVIEW --------------------

The work takes the view that many services are made available independently and will likely include several near or perfect duplicates - all-be-it using different terminology. Starting from this point of view, one can use a data mining approach to identify sets of services which can be considered equivalent. If services are made available through a more formal procedure, it might be argued that the classification into equivalence sets will happen naturally.

In either case, given a set of operations classed as equivalents there remains the issue of how best to choose a particular service on the basis of non-functional requirements. The authors have developed in previous work a model of a service pool - which effectively represents an equivalence class. The service pool is a virtual service, whose interface is a sort of average of those of the pool members it represents, and which redirects calls to specific members on the basis of the non-functional requirements. In effect the virtual service acts much like a broker. This paper claims two contributions: the first being a method for computing the service pool given a set of services; the second being an efficient method for selecting a service according to non-functional requirements.

The technique chosen for computing the service pool is based on woogle, and appears to be a very slight modification of that technique. Essentially, they claim that the original woogle algorithm (repeated steps of aglomaterive clustering with cluster-splitting) tends to yield modest precision - at around 60-70% for operation matching, so they propose to filter results according to one of the UDDI fields (domainkey). This filtering technique appears to be quite orthogonal to the woogle procedure. They say this increases precision to about 85%, but don't give any contextual details, and don't appear to present any large scale implementation results, e.g. using third party public services, which might be interesting.

The contribution to the QoS matching phase is the presentation of an algorithm which is claimed to have polynomial cost for the selector. The idea is, given a set of QoS metrics, to pre-compute orderings of services so that selection at call-time is faciltated. This approach is valid only to the extent that QoS metrics tend to remain static for a significant interval; an assumption stated by the authors. In practice it seems that the approach could be useful in some scenarios but not in others. For instance, one of the metrics chosen is, as expected, response time. It could be argued that if such a metric is to be useful to the user, it should include transfer time, which seems very unlikely to be a very static cost. Furthermore, in a real environment it seems reasonable to assume that services will be accessed by multiple concurrent users; surely then the degree of concurrent access must affect the response time of any individual request. This could happen either through slowing execution of concurrent executions or through imposing an arbitrary delay through queueing of requests. It seems that in a ve! ry dynamic environment, the pre-computation method could turn out to be more expensive than an alternative which deferred all computation to call-time, because it is investing in a global pre-computation on the assumption of being able to amortize this investment over a period of time. Invested computation will also be wasted if the pre-computed orderings are not used sufficiently between changes in QoS metrics. On the other hand if transfer costs can be assumed to be negligible, for instance since the operation costs are very large, and concurrent access to services (and the machines hosting them) is precluded, and requests are not biassed to a small subset of the overall set of operation classes, then the overall approach seems reasonable. The authors do include some experiments which report the variation in cost of their search algorithm with pool size, but these seem to give limited support to the claim for low cost; ignoring the point corresponding to pool size of 1! 9 (which seems unexpectedly high), the cost appears to increase by a f actor of two as the pool size is increased by 1. I presume this is a feature of the implementation being preliminary and therefore not so well tuned.

Overall, Id suggest the work is not making a very great contribution in the pool setup phase, and is making marginal contribution in the QoS matching phase, so in its current form, should not be accepted for www2007.

While a woogle type approach and any extension to it is interesting, the work on QoS matching actualy appears currently to be basically orthogonal. It is thus possible to separate the two parts of the paper. As well as gaining some extra space to describe the QoS matching or permitting a submission of a shorter paper, the QoS matching component can be seen as having a greater applicability, being relevant to any scenario where a broker selects between similar services. I'd encourage the authors to consider reworking the paper along these lines - and to include more recent experimental results.

-------------------- review 2 --------------------


----------------------- REVIEW --------------------

QoS-aware service discovery is a hot topic in Web services. This paper introduces a service pool based service discovery and request dissemination procedure, a formalized QoS model from consumer's perspective is also proposed with a QoS-aware discovery algorithm to select the most adequate provider. The idea is interesting, and the result is evaluated by extensive experiment. However, the paper can be improved in some distance:

First, from Fig. 4, we can see that the efficiency of this approach may be lower when there are more candidate services in the pool. This may introduce serious scalability problem to the proposed mechanism. Actually there are some other approaches to solve the related problem, such as QoS-aware UDDI and distributed UDDI. These may be help to improve the performance of the mechanism proposed in the paper.

Second, The QoS model proposed in the paper is not a new idea. We know that there are many QoS dimensions in Web service QoS domain. But only several QoS dimensions are mentioned in this paper, such as successability, response time, price cost and reputation. Is the model available to other QoS dimensions? A further survey on Qos model may be useful to the author.

Third, the experiment shows the availability of the proposed mechanism. It抯 fine, but would be much better to design extensive simulations or experiment to compare with existing service discovery method to show more value of the service pool based service discovery.

In addition, the writing quality can still be improved. Some sentences are not easy to understand. Also, there are some typos in the paper. For example, Paragraph 3 in section 3.2 : " two steps " -> " three steps ".

I suggest the authors consider the above issues to improve the paper.

-------------------- review 3 --------------------


----------------------- REVIEW --------------------

This paper describes an interesting scheme for grouping services and providing mechanisms so that consumers?QoS requirements can be met by automatically selecting the service that best matches their requirements.

It appears to assume that all web services have an rpc-like interface.. but most do, so that need not invalidate the work. However the scope should be made clear.

The contributions are in the area of: a) how to group services b) how to select a service that meets the users requirements

I found it a very interesting paper, but I had concerns over whether it could be applied in practise. For example: - in what circumstances would a ws consumer choose a service by keywords or a directory of services? How would they be confident that it would do what they wanted? As web services are designed for machine to machine interaction rather than end user to service interaction then is this method of finding services appropriate?

The evaluation is a little basic and doesn抰 address these issues (e.g. by analysing in what proportion of the time the selected service has the desired semantics).

However, the method of selecting a service based on QoS requirements is interesting, and independent of the way in which services are grouped (for example they could be grouped manually by a user or trusted 3rd party).

Therefore, overall I thought this was thought provoking and interesting even though it raises issues of practicality. As a result I have decided to recommend it抯 acceptance.