CAP F07/Ethics Team 2/Paper

Maintainers: Tony Waldschmidt, John Giannini, Will Voorhees;

Introduction to St. Olaf Cluster Computing Program
In the fall of 2003 discussion began regarding the possibility of a St. Olaf cluster computing facility and when a grant was written for the Howard Hughes Medical Institute (HHMI) it included a brief mention of a funding request for "acquisition of a Beowulf cluster, servers, and workstations". In 2004 St. Olaf's grant was approved with a total of 1.3 million dollars, $6,000 of which was alloted for computing resources. From these humble beginnings the St. Olaf cluster computing project has grown to tens of thousands of dollars and over a dozen different research projects. The purpose of this paper is to analyze roles of scientific and ethical conduct in relationship to the St. Olaf Cluster Computing Program.

STS:

What role does the concept of scientific conduct play in the cluster?
The recent addition of the helios and castaway clusters to the computer science department has ushered in a new level of interdisciplinary cooperation with the sciences in realm of research. Traditional scientific research has a well-developed ethical system for measuring scientific misconduct and one question we would like to explore is the ethical ramifications of introducing this type of research onto the cluster.

Inevitably in scientific inquiry, investigators are to some extent on their honor not to falsify results. It was the unanimous consensus among the faculty we talked to that such fraud is always a possibility for an unscrupulous scientist. As Professor Freedberg noted, even if records are maintained with scrupulous care (in his case careful cataloging and even video recording of experiments), there is always the possibility that the data presented to the public has undergone selective omissions. No matter how hard he might try to put his methods beyond question his findings must always rest, at least to some extent, on his credibility. [4]

There seems to be no reason to think that research conducted on the cluster will be any different in this regard. The cluster is used like any other scientific instrument in that it delivers its results to the researcher to interpret and relay, perhaps selectively.

One factor which might make researchers more hesitant to falsify computer-based findings than they would be others is the ease with which such findings might be replicated. Though it is a goal of scientific literature to provide enough detail on experimental method that the experiment can be duplicated and confirmed, it may be difficult to actually do in many cases, as it might require significant setup, equipment, technical expertise, and time. Further, in many biological cases, it is always possible that the subjects in the original study differ from those in the repeated study in some way (the second study does not observe the same forest for the same years, or dissect the same turtles).

In contrast, it could be much easier to implement a computer-based project in a new environment. It is customary to provide the code used in research papers and thus checking the findings might be as simple as plugging in the code. Also, this part of the process seems like it should not be subject to the variation mentioned above: given the same code and input, a computer should always generate the same result. Thus, in so far as findings are based on replicatable computations, and the methods employed are accurately reported, computer based research may make researchers more easily answerable to the broader scientific community.

This raises related question about the accuracy of computer models. Computer research almost inevitably lends itself to the use of models due to its abstraction from any actual data gathering. In our limited observation, model based simulations have made up a significant portion of cluster use; and it seems likely that they will continue to. Professor Schade suggests that this tendency itself may be problematic to good scientific procedure. To his way of looking at things, it is dangerous to use models which produce findings that can’t be checked, at least partially, against actual empirical observation. If the power of computers in model-based inquiry encourages such abstraction, Schade feels that the scientific results could be dubious if not carefully checked. [5]

Professor Freedberg is significantly more optimistic about the value of models which must proceed in the absence of hard data. He points out that some inquiry is, by its very nature, excluded from currently achievable empirical verification (for example, models dealing with population dynamics over hundreds of years). While models should be based on the known facts, and verified through observation whenever possible, he feels that they can still be very useful in the regions beyond such fact checking. [4]

On the other hand, he was even more emphatic on a point of concern which he and Doctor Shade share: that of the dependence of a model on a scientist. Even if a researcher does not selectively report results, computer models are by nature artifacts in that they are hypothetical constructions of scientists. As such, they are somewhat plastic, and may reflect the bias of their designer in their very structure. Here again, the scientific community is relied on to look carefully at the code used and question assumptions inherent in it.

How do undergraduate students acting as research assistants change the STS, and how does that impact faculty and student prioritization of goals (e.g. educational vs. research goals)?
Computer science classes such as the senior capstone often put students in the position of doing cluster-related work for an ongoing faculty research project. Ideally, students achieve both educational and scientific ends through these research projects. However, a potential concern is that students acting as research assistants because of classroom obligations may lead to some conflict of interest, specifically for the faculty. What effects is this arrangement likely to have on the socio-technical system already in-place? Is it possible that what best furthers the research and development goals is not necessarily what is most educationally useful? If this were the case, mightn’t it tempt faculty members to put research ends above educational ones?

One of the effects this has on the socio-technical system is that the software and hardware being used by a given research program may now include elements with which the manager of the research program does not interact directly. Students have become intermediaries between devices and researchers. For instance, whereas previously a researcher might have run Stata on her own computer if they were to use it, now a faculty researcher can delegate the parallelization and running of the software to a student researcher. This may be mitigated by the development of web-based interfaces, but, so far, students have been relied upon to develop these as well. While it isn't out of the ordinary for such an intermediary to exist (in the form of colleagues or professional technicians, for instance) the feature that is remarkable in this situation is that the go-betweens are non-professional and non-voluntary (non-voluntary in the sense that such activity is now essential to the St. Olaf C.S. major).

One salient category of our STS analysis regards the people involved with the cluster. While the use of students as research aids doesn't necessarily change this group significantly, it does bring about important changes in roles and dependencies. For one, it extends the group of people involved in a research project to include C.S. students, teachers, and support staff who manage the courses involved and not merely those who operate the cluster itself. Also, potentially, it involves IIT staff and policies in the situation, since they operate the network infrastructure and provide hardware. (As well as involving many more people, this potentially puts those people in a position of authority, as they manage use of the cluster. The analysis of procedures and policies related to this issue is covered in the third section of our analysis.)

These changes have been fairly minor so far, especially as few, if any, of the research projects currently using the cluster actually rely on it centrally in their endeavors. However, this may change as current development aims to make cluster use central to as many research programs as possible, which in turn would likely make these issues much more significant, and the same can be said of the concerns which follow from them.

The way in which these changes make research projects dependent on students whose goal is solely educational was our primary concern here. As mentioned above, we speculated that the aims of research could, at times, be at odds with those of education, and thus that faculty might be tempted to prioritize their individual research ahead of the goal of educating students.

Some might not agree with the idea that research and educational ends are at odds. Professor Schade indicated that he thought research involvement was of inherent educational value, and didn't envision a scenario where the two ends might diverge.

Although the ideal of research as a educational boon is attractive, we can envision scenarios where education and research do not entirely lie along the same path. In a scenario suggested by professor Freedberg, a researcher faced with a time deficit must decide between haveing a student finish the remaining five-percent of their research through familiar and monotonous data gathering, or giving the student a full conceptual understanding of the subject through some new exploration. He not only saw the potential for such a situation to arise, but suggested that the choice would be a tricky one for most.

Professor Freedberg did point out, however, that such extreme cases aren't likely to be common, and that, for the most part, the ends of both teacher and researcher can be achieved together. In his own case, he said that if at any point he feels a student's educational experience has been overshadowed he makes certain to err in the other direction on subsequent occasions.

In what situations is use of the cluster justified? In what situations is it not? Who makes these decisions?
The third point of investigation in our ethical analysis sought to address the ethical aspects of cluster use and management. Specifically we sought to explore when cluster use is or is not justified and who is involved in assessing and regulating these decisions and allowing different projects to go forward.

To develop a framework and basis of comparison, we decided to conduct our investigation by researching how external educational institutions manage some of these same issues. Based primarily on public availability of cluster policies (linked below), we selected three institutions to compare: Harvey Mudd College in Claremont, California; San Francisco State University; and Tufts University in Boston, Massachusetts. As mentioned previously, our primary concerns in developing an analytical comparison of each schools cluster policy were appropriate use, and authority structures involved in this decision.

Unfortunately, all three schools were extremely vague in terms of what constituted appropriate cluster usage. Although we can only speculate as to the thinking behind the lack of detailed project specifications, perhaps we can extrapolate some of the reasons for this lack of transparency. One primary reason that will be discussed in detail in the next section is the application barrier. Because each cluster's controlling authority structure explicitly forbids any usage not outlined by the application process, this authority structure is not dealing with the issues of appropriate use publicly but rather during the internal selection process. Another plausible reason contributing to the lack of public documentation or discussion on appropriate use on each institutions website is the generalization of regulation to larger authority structures. All three institutions held this feature in common in that each cluster policy contains a reference to a higher authority whose regulations on use must also be abided. The chain of authority structures was as follows for the three institutions:  SFSU : Center for Computing Life Science -> IT resources acceptable use -> University Student Code of Conduct -> Federal and State Law [2] Harvey Mudd : Math Department -> Harvey Mudd College Standards of Conduct -> Federal and State Law [1] Tufts : general IT policy -> standard university ethical and legal conduct -> Federal and State Law [2]  This generalization is logical and probably necessary. However, simply referring the user up an authority chain obfuscates exactly what constitutes proper use and reduces the likelihood of full comprehension of these regulations on the part of the user. As a final note, all three institutions did provide this minimal guidance of allowing only approved work that fell within legal guidelines of larger authority structures. Harvey Mudd and SFSU also forbid commercial work on their respective clusters [1][2] and SFSU added a clause forbidding political advocacy [2].

The second key aspect of cluster policy we sought to examine was authority structures in relation to control of cluster access. One interesting commonality across all three institutions is the immediate encounter between the end user and bureaucracy with it's structured restrictions. Specifically all three immediately require the user to fill out an application in order to carry out research on the cluster. Some of the institutions had higher barriers than others to gain access even to this preliminary step. SFSU for example only allows principle investigators (PI) to make requests on a per-project basis. Students seeking to utilize the cluster must do so through a PI [2]. Harvey Mudd retains a similar faculty-standing requirement but adds another bureaucratic hoop. The faculty must apply for a Mathematics user account and upon approval may apply for use of the AMBER cluster by detailing the desired project [1]. Tufts stands in contrast to these two by allowing any student, faculty, or staff to apply for cluster use. [3]

These bureaucratic barriers to cluster usage beget the question of origin: who is imposing these limitations? This appears to differ on an institutional basis and unsurprisingly is typically dictated by those who run the cluster. At SFSU, the Center for Computing Life Sciences operates the cluster and is in charge of receiving applications and approving appropriate projects [2]. The authority structure administering and regulating access to the Harvey Mudd cluster took the form of an educational department (Mathematics and Computer Science) rather than a separate center for computing [1]. Tufts was also unique in that the cluster was a resource integrated into the oversight of the University Information Technology group (IIT equivalent) and was consequently regulated by this IT department. [3]

In addition to deciding which types of usage are appropriate from an academic standpoint it is also important to discuss if the creation and use of a Beowulf cluster is indeed the best ethical use of our resources. The purchase of the hardware for the Beowulf cluster cost tens of thousands of dollars and its operation has consumed countless hours of very intelligent people's time. How than can we justify its purchase? Furthermore, in addition to the opportunity cost of the labor and money invested in the beowulf project, we must also consider the environmental impact of it's production and operation.

Each machine consumes 188.3W during peack usage according to the | sun server power calculator. with 16 machines that represents a total of 3 012.8 Watts. Assuming that one third of that power comes from the wind turbine, and the remaining two thirds comes from sources typical to the state of Minnesota, we can use the governmental | Emission Factor to produce an estimate about the carbon footprint. Using this data we can estimate that when the cluster is running at full capacity it produces about 4.56 lbs of carbon per hour, or 20 tons of carbon per year. To give some perspective, it is estimate that a typical North American produces about five tons of carbon per year. In addition to the ongoing energy consumption associated with the operation of the cluster, we must also consider the environmental impact of producing these machines, this concept is called embodied energy. Calculating the embodied energy of the the machines is exceedingly difficult due to the wide composition of materials that go into each machine. We can however assume that this cost is quite high, the materials that go into a computer are often toxic and take special care to dispose of. The toxic nature of computers has created a push for Restriction of Hazardous Substances (RoHS) compliant machines, a standard which has been mandated in the European Union. Fortunately our Sun X2100s are indeed RoHS compliant.

Offsetting all of these costs are the benefits of the use of the cluster, research using the cluster has spanned many different disciplines on campus and produced plenty of valuable research. While it is impossible to objectively weight the value of the research verse the costs of the Beowulf clusters production and operation, hopefully this analysis will help us remain mindful of the costs incurred and maximize the returns that it produces.