Home » Research Ideas
Category Archives: Research Ideas
Challenges of Capturing System Activity
A key aspect of the work I am doing for Indaleko is to “capture system activity” so that it can be used to form “activity contexts” that can then be used to inform the process of finding relevant information. As part of that, I have been working through the work of Daniela Vianna. While I have high-level descriptions of the information she collected and used, I need to reconstruct this. She collects data from a variety of sources. The most common source of information comes from web APIs to services such as Google and Facebook. In addition, she also uses file system activity information.
Since my background is file systems, I decided to start on the file system activity front first. Given that I’ve been working with Windows for three decades now, I decided to leverage my understanding of Windows file systems to collect such information. One nice feature of the NTFS file system on Windows is its support for a form of activity log known as the “USN Journal.” Of course, one of my handicaps is that I am used to using the native operating system API, not the libraries that are implemented on top of it. This is because when building file systems on Windows I have always been interested in testing the full kernel file systems interface. While there are a few specific features that cannot be exercised with just applications, there are still a number of interfaces that cannot be tested using the typical Win32 API that can be tested using the native API. In recent years the number of features that have been hidden from the Win32 API has continued to decrease, which has diminished the need to use the native API. I just haven’t had any strong need to learn the Win32 API – why start now?
I decided the model I want to use is a service that pulls data from the USN journal and converts it into a format suitable for storing in a MongoDB database. I decided to go with Mongo because that is what Vianna used for her work. The choice at this point is somewhat arbitrary but MongoDB makes sense because it tends to work well with semi-structured data, which is what I will be handling.
Similarly, I decided that I’d write my service for pulling USN Journal data from the NTFS file system(s) in C# since I have written some C# in the past, it makes doing some of the higher level tasks I have much easier, and is well-supported on Windows. I have made my repository public though I may restructure and/or rename it at some point (currently I call it CSharpToNativeTest because I was trying to invoke the native API as unmanaged code from C#). The most common approach to this is to utilize a specific mechanism (the “PInvoke” mechanism) but after a bit of trial-and-error I decided I wanted something that would be easier for me to debug, so instead of pulling the native routine directly from ntdll.dll I load it from my own DLL (written in C) and that then invokes the real native call. This allows me to see how data is being marshaled and delivered to the C language wrapper. I also tried to make the native API “more C# friendly.” I am sure it could be more efficient, but I wanted to support a model that I could extend and hopefully it will be easier to make it more efficient should that prove necessary.
One thing I did was to script the conversion of all the status values in ntstatus.h into a big C# enum type. The benefit of this is that when debugging I can automatically see the mnemonic name of the status code as well as its numeric value. I then decided to provide the layer needed to map the various volume names used on Windows around, with device names, device IDs, and symbolic links (drive letters) that can be mapped. While I have not yet added it, I wrote things so that it should be fairly straight-forward to add a background thread which wakes up when devices arrive or disappear. As I have noted before “naming is hard.” This is just one more example of the flexibility and challenges with aliasing and naming.
Finally, I turned my attention to the USN journal. I found some packages for decoding USN journal entries; most were written to parse the data from the drive, while a few managed dynamic access. Since I want this to be a service that monitors the USN journal and keeps adding information into the database, I decided to write C# code to use the API for retrieving that information. At this point, what I have is the ability to scan all the volumes on the machine – even if they do not have drive letters – and query them to see if they support a USN journal. I do this properly – I query the file system attributes (using the NtQueryVolumeInformationFile native API) and check if the bit showing USN journal support is marked. I do not use the file system name, an approach I’ve always considered to be a hack, especially since I have been in the habit of writing file systems that support NTFS features, including named data streams, extended attributes, and object IDs. In fact, the ReFS file system on Windows also supports USN journals, so I’m not just being my usual pedantic developer self in this instance.
At this point, I am able to identify volumes that support USN journals, open them and find out if USN is turned on (it is by default on the system volume, which is almost always the “C:” drive, though I enjoy watching things break when I configure a system to use some other drive letter.) I then extract the information and convert it to in-memory records. At the moment I just have it wait a few seconds and pull the newest records, but my plan is to evolve this into a service that I can run and it can keep pulling data and pushing it into my MongoDB instance.
At this point, I realized I do not really know that much about MongoDB so I have decided to start learning a bit more about it. Of course, I don’t want to be a MongoDB expert, so I also have been looking more carefully at Daniela Vianna’s work, trying to figure out what her data might have looked like and think about how I’m going to merge what she did into what I am doing. This is actually exciting because it means I’m starting to think of what we can do with this additional information.
This afternoon I had a great conversation with one of my PhD supervisors about this and she was making a couple of suggestions about ways to consume this data. That she was suggesting things I’d also added to my list was encouraging. What are we thinking:
- We can consider using “learned index structures” as we begin to build up data sets.
- We can use techniques such as Google BERT to facilitate dealing with the API data that Vianna’s work used. I pointed out that the challenges of APIs that Vianna pointed out are similar to languages: they have meaning and those meanings can be expressed in multiple ways.
- The need for being able to efficiently find things is growing rapidly. She was explaining some work that indicates our rate of data growth is outstripping our silicon capabilities. In other words, there is a point at which “brute force search” becomes impractical. I liked this because it suggests what we are seeing with our own personal data is a leading indicator of the larger problem. This idea of storing the meta-data independent of the data is a natural one in a world where the raw information is too abundant for us to just go looking for an item of interest.
So, my work continues, mostly mundane and boring, but there are some useful observations even at this early stage. Now to figure out what I want the data in my database to look like and start storing information there. Then I can go figure out what I did right, what I did wrong, and how to improve things.
Aside: one interesting aspect of the BERT work was their discussion of “transducers.” This reminded me of Gifford’s Semantic File System work, where he used transducers to suck out semantic information from existing files.
Brainiattic: Remember more with your own Metaverse enhanced brain attic
I recently described the idea of “activity context” and suggested that providing this new type of information about data (meta-data) to applications would permit improve important tasks such as finding. My examining committee challenged me to think about what I would do if my proposed service – Indaleko – already existed today.
This is the second idea that I decided to propose on my blog. My goal is to find how activity context can be used to provide enhanced functionality. My first idea was fairly mundane: how can we improve the “file browsing” experience in a fashion that focuses on content and similarity by combining prior work with the additional insight provided by activity context.
My initial motivation for this second idea was motivated by my mental image of a personal library but I note that there’s a more general model here: displaying digital objects as something familiar. When I recently described this library instantiation of my brain attic the person said “but I don’t think of digital objects as being big enough to be books.” To address this point: I agree, another person’s mental model for how they want to represent digital data in a virtual world need not match my model. That’s one of the benefits of virtual worlds – we can represent things in forms that are not constrained by what things must be in the real world.
In my recent post about file browsers I discussed Focus, an alternative “table top” browser for making data accessible. One reason I liked Focus is that the authors observed how hierarchical organization does not work in this interface. They also show how the interface is useful and thus it is a concrete argument as to at least one limitation of the hierarchical file/folder browser model. Another important aspect of the Focus work was their observation that a benefit of the table top interface is it permits different users to organize information in their own way. A benefit of a virtual “library” is that the same data can be presented to different users in ways that are comfortable to them.
Of course, the “Metaverse” is still an emerging set of ideas. In a recent article about Second Life Philip Rosedale points out that existing advertising driven models don’t work well. This begs the question – what does work well?
My idea is that by having a richer set of environmental information available, it will be easier to construct virtual models that we can use to find information. Vannevar Bush had Memex, his extended memory tool. This idea turns out to be surprisingly ancient in origin, from a time before printing when most information was remembered. I was discussing this with a fellow researcher and he suggested this is like Sherlock Holmes’ Mind Palace. This led me to the model of a “brain attic” and I realized that this is similar to my model of a “personal virtual library.”
The Sherlock Holmes article has a brilliant quotation from Maria Konnikova: “The key insight from the brain attic is that you’re only going to be able to remember something, and you can only really say you know it, if you can access it when you need it,”
This resonates with my goal of improving finding, because improving finding improves access when you need it.
Thus, I decided to call this mental model “Braniattic.” It is certainly more general than my original mental model of a “personal virtual library,” yet I am also permitted to have my mental model of my pertinent digital objects being projected as books. I could then ask my personal digital librarian to show me works related to specific musical bands, or particular weather. As our virtual worlds become more capable – more like the holodeck of Star Trek – I can envision having control of the ambient room temperature and even the production of familiar smells. While our smart thermostats are now capturing the ambient room temperature and humidity level and we can query online sources for external temperatures, we don’t actively use that information to inform our finding activities, despite the reality is that human brains do recall such things; “it was cold out,” “I was listening to Beethovan,” or “I was sick that day.”
Thus, having additional contextual information can be used at least to improve finding by enabling your “brain attic.” I suspect that, once activity context is available we will find additional ways to use it in constructing some of our personal metaverse environments.
Better Finding: Combine Semantic and Associative Context with Indaleko
Last month I presented my thesis proposal to my PhD committee. My proposal doesn’t mean that I am done, rather it means that I have more clearly identified what I intend to make the focus of my final research.
It has certainly taken longer to get to this point than I had anticipated. Part of the challenge is that there is quite a lot of work that has been done previously around search and semantic context. Very recent work by Daniela Vianna relates to the use of “personal digital traces” to augment search. It was Dr. Vianna’s work that provided a solid theoretical basis for my own proposed work.
Our computer systems collect quite an array of information, not only about us but also about the environment in which we work.
In 1945 Vannevar Bush described the challenges to humans of finding things in a codified system of records. His observations continue to be insightful more than 75 years later:
Our ineptitude in getting at the record is largely caused by the artificiality of systems of indexing. When data of any sort are placed in storage, they are filed alphabetically or numerically, and information is found (when it is) by tracing it down from subclass to subclass. It can be in only one place, unless duplicates are used; one has to have rules as to which path will locate it, and the rules are cumbersome. Having found one item, moreover, one has to emerge from the system and re-enter on a new path.
The human mind does not work that way. It operates by association. With one item in its grasp, it snaps instantly to the next that is suggested by the association of thoughts, in accordance with some intricate web of trails carried by the cells of the brain. It has other characteristics, of course; trails that are not frequently followed are prone to fade, items are not fully permanent, memory is transitory. Yet the speed of action, the intricacy of trails, the detail of mental pictures, is awe-inspiring beyond all else in nature.
I find myself returning to Bush’s observations. Those observations have led me to ask if it is possible for us to build systems that get us closer to this ideal?
My thesis is that collecting, storing, and disseminating information about the environment in which digital objects are being used provides us with new context that enables better finding.
So, my proposal is about how to collect, store, and disseminate this type of external contextual information. I envision combining this with existing data sources and indexing mechanisms to allow capturing activity context in which digital objects are used by humans. A systems level service that can do this will then enable a broad range of applications to exploit this information to reconstruct context that is helpful to human users. Over my next several blog posts I will describe some ideas that I have with what I envision being possible with this new service.
The title of my proposal is: Indaleko: Using System Activity Context to Improve Finding. One of the key ideas from this is the idea that we can collect information the computer might not find particularly relevant but the human user will. This could be something as simple as the ambient noise in the user’s background (“what music are you listening to?” or “Is your dog barking in the background”) or environmental events (“it is raining”) or even personal events (“my heart rate was elevated” or “I just bought a new yoga mat”). Humans associate things together – not in the same way, nor the same specific elements – using a variety of contextual mechanisms. My objective is to enable capturing data that we can then use to replicate this “associative thinking” that helps humans.
Ultimately, such a system will help human users find connections between objects. My focus is on storage because that is my background: in essence, I am interested in how the computer can extend human memory without losing the amazing flexibility of that memory to connect seemingly unrelated “things” together.
In my next several posts I will explore potential uses for Indaleko.
Laundry Baskets: The New File System Namespace Model
While I’ve been quiet about what I’ve been doing research-wise, I have been making forward progress. Only recently have ideas been converging towards a concrete thesis and the corresponding research questions that I need to explore as part of verifying my thesis.
I received an interesting article today showing that my research is far more relevant than I’d considered: “FILE NOT FOUND“. The article describes that the predominant organizational scheme for “Gen Z” students is the “Laundry Basket” in which all of their files are placed. This is coming as a surprise to people who have been trained in the ways of the hierarchical folder metaphor.
While going through older work, I have found it is intriguing that early researchers did not see the hierarchical design as being the pinnacle of design; rather they saw it as a stop-gap measure on the way to richer models. Researchers have explored richer models. Jeff Mogul, now at Google Research, did his PhD thesis around various ideas about improving file organization. Eno Thereska, now at Amazon, wrote an intriguing paper while at Microsoft Research entitled “Beyond file systems: understanding the nature of places where people store their data” in which he and his team pointed out that cloud storage was creating a tension between file systems and cloud storage. The article from the Verge that prompted me to write this post logically makes sense in the context of what Thereska was saying back in 2014.
The challenge is to figure out what comes instead. Two summers ago I was fortunate enough to have a very talented young intern working with me for a couple months and during that time one of the interesting things he built was a tool that viewed files as a graph rather than a tree. The focus was always at the center, but then it would be surrounded by related files. Pick one of those files and it became the central focus, with a breadcrumb trail showing how you got there but also showing other related files.
The relationships we used were fairly simple and extracted from existing file meta-data. What was actually quite fascinating about it though was that we constructed it to tie two disjoint storage locations (his local laptop and his Google Drive) together into a single namespace. It was really an electrifying demonstration and I have been working to figure out how to enable that more fully – what we had was a mock-up, with static information, but the visualization aspects of “navigating” through files was quite powerful.
I have been writing my thesis proposal, and as part of that I’ve been working through and identifying key work that has already been done. Of course my goal is to build on top of this prior work and while I have identified ways of doing this, I also see that to be truly effective it should use as much of the prior work as possible. The idea of not having directories is a surprisingly powerful one. What I hadn’t heard previously was the idea of considering it to be a “laundry basket” yet the metaphor is quite apt. Thus, the question is how to enable building tools to find the specific thing you want from the basket as quickly as possible.
For example, the author of the Verge article observed: “More broadly, directory structure connotes physical placement — the idea that a file stored on a computer is located somewhere on that computer, in a specific and discrete location.” Here is what I recently wrote in an early draft of my thesis proposal: “This work proposes to develop a model to separate naming from location, which enables the construction of dynamic cross-silo human usable name-spaces and show how that model extends the utility of computer storage to better meet the needs of human users.”
Naming tied to location is broken, at least for human users. Oh, sure, we need to keep track of where something is stored to actually retrieve the contents, but there is absolutely no reason that we need to embed that within the name we use to find that file. One reason for this is that we often choose the location due to external factors. For example, we might use cloud storage for sharing specific content with others. People that work with large data sets often use storage locations that are tuned to the needs of that particular data set. There is, however, no reason why you should store the Excel spreadsheet or Python notebook that you used to analyze that data in the same location. Right now, with hierarchical names, you need to do so in order to put them into the “right directory” with each other.
That’s just broken.
However, it’s also broken to expect human users to do the grunt work here. The reason Gen Z is using a “laundry basket” is because it doesn’t require any effort on their part to put something into that basket. The work then becomes when they need to find a particular item.
This isn’t a new idea. Vannevar Bush described this idea in 1945:
“Consider a future device for individual use, which is a sort of
mechanized private file and library. It needs a name, and, to coin
one at random, ”memex” will do. A memex is a device in which
an individual stores all his books, records, and communications,
and which is mechanized so that it may be consulted with exceeding
speed and flexibility. It is an enlarged intimate supplement to his
memory.”
He also did a good job of explaining why indexing (the basis of hierarchical file systems) was broken:
“Our ineptitude in getting at the record is largely caused by the artificiality of systems of indexing. When data of any sort are placed in storage, they are filed alphabetically or numerically, and information is found (when it is) by tracing it down from subclass to subclass. It can be in only one place, unless duplicates are used; one has to have rules as to which path will locate it, and the rules are cumbersome. Having found one item, moreover, one has to emerge from the system and re-enter on a new path.
“The human mind does not work that way. It operates by association. With one item in its grasp, it snaps instantly to the next that is suggested by the association of thoughts, in accordance with some intricate web of trails carried by the cells of the brain. It has other characteristics, of course; trails that are not frequently followed are prone to fade, items are not fully permanent, memory is transitory. Yet the speed of action, the intricacy of trails, the detail of mental
pictures, is awe-inspiring beyond all else in nature.”
We knew it was broken in 1945. What we’ve been doing since then is using what we’ve been given and making it work as best we can. We knew it was broken. Seltzer wrote “Hierarchical File Systems Are Dead” back in 2009. Yet, that’s what computers still serve up as our primary interface.
The question then is what the right primary interface is. Of course, while I find that interesting I work with computer systems and I am equally concerned about how we can build in better support, using the vast amount of data that we have in modern computer systems, to construct better tools for navigating through the laundry basket to find the correct thing.
How I think that should be done will have to wait for another post, since that’s the point of my thesis proposal.
Where has the time gone?
It’s been more than a year since I last posted; it’s not that I haven’t been busy, but rather that I’ve been trying to do too many things and have been (more slowly than I’d like) cutting back on some of my activities.
Still, I miss using this as a (one way) discussion about my own work. In the past year I’ve managed to publish one new (short) paper, though the amount of work that I put into it was substantial (it was just published in Computer Architecture Letters). This short article (letter) journal normally provides at most one revise and resubmit opportunity, but they gave me two such opportunities, then accepted the paper, albeit begrudgingly over the objections of Reviewer # 2 (who agreed to accept it, but didn’t change their comments).
Despite the lack of clear publications to demonstrate forward progress, I’ve been working on a couple of projects to push them along. Both were presented, in some form, at Eurosys as posters.
Since I got back from a three month stint at Microsoft Research (in the UK) I’ve been working on one of those, evolving the idea of kernel bypasses and really analyzing why we keep doing these things; this time through the lens of building user mode file systems. I really should write more about it, since that’s on the drawing board for submission this fall.
The second idea is one that stemmed from my attendance at SOSP 2019. There were three papers that spoke directly to file systems:
- File Systems Unfit as Distributed Storage Backends
- Performance and Protection in the ZoFS User-space NVM File System
- SplitFS: Reducing Software Overhead in File Systems for Persistent Memory
Each of these had important insights into the crossover between file systems and persistent memory. One of the struggles I had with that short paper was explaining to people “why file systems are necessary for using persistent memory”. I was still able to capture some of what I’d learned, but a fair bit of it was sacrificed to adding background information.
One key observation was around the size of memory pages and their impact on performance; it convinced me that we’d benefit from using ever larger page sizes for PMEM. Some of this is because persistent memory is, well, persistent and thus we don’t need to “load the contents from storage”. Instead, it is storage. So, we’re off testing out some ideas in this area to see if we can contribute some additional insight.
The other area – the one that I have been ignoring too long – is the thesis of this PhD work in the first place. Part of the challenge is to reduce the problem down to something that is tractable and can be finished in a reasonable amount of time.
One of the questions (and the one I wanted to explore when I started writing this) is a rather famous article from 1945 entitled As We May Think. Vannevar Bush described something quite understandable, yet we have not achieve this, though we have been trying – one could argue that hypertext stems from these ideas, but I would argue that hypertext links are a pale imitation of the rich assistive model Bush lays out when he describes the Memex.
Thus, to the question, which I will reserve for another day: why have we not achieved this yet? What prevents us from having this, or something better, and how can I move us towards this goal?
I suspect, but am not certain, that one culprit may be the fact we decided to stick with an existing and well-understood model of organization:
360° Semantic File System: Augmented Directory Navigation for Nonhierarchical Retrieval of Files
360° Semantic File System: Augmented Directory Navigation for Nonhierarchical Retrieval of Files, Syed Rahman Mashwani and Shah Kusro, in IEEE Access, January 29, 2019.
This paper makes some interesting observations that resonate with my own research observations, though I will end up arguing (in a future blog post) that they don’t go far enough. But they do a good job of laying out the problem and why some solutions do not work. One of the common themes I have heard when discussing my own work is an insistence there really isn’t a problem, though usually a longer conversation ends up with us agreeing things could be done better.
The abstract is a bit long, but clearly describes the essence of the paper:
The organization of files in any desktop computer has been an issue since their inception. The file systems that are available today organize files in a strict hierarchy that facilitates their retrieval either through navigation, clicking directories and sub-directories, in a tree-like structure or by searching (which allows for finding of the desired files using a search tool). Research studies show that the users rarely (4% – 15%) use the latter approach, thus leaving navigation as the main mechanism for retrieving files.
Paper Abstract
However, navigation does not allow a user to retrieve files nonhierarchically, which makes it limited in terms of time, human effort, and cognitive overload. To mitigate this issue, several semantic file systems (SFSs) have been periodically proposed that have made the nonhierarchical navigation of files possible by exploiting some basic semantics but no more than that. None of these systems consider aspects such as time, location, file movement, content similarity, and territory together with learning from user file retrieval behaviors in identifying the desired file and accessing it in less time and with minimum human and cognitive efforts.
Moreover, most of the available SFSs replace the existing le system metaphor, which is normally not acceptable to users. To mitigate these issues, this research paper proposes 360 SFS that exploits the SFS ontology to capture all the possible relevant file metadata and learns from user browsing behaviors to semantically retrieve the desired files both easily and timely. Based on user studies, the evaluation results show that the proposed 360 SFS outperforms the existing traditional directory navigation and recently open files.
Of course, the problem existed before the appearance of the desktop computer: the original UNIX contained the permuted index server, which suggests to me that even in 1973 people were struggling to find things. What I do find interesting is the observation that people really do not like search – I recently described this insightful (to me at least) observation. Here it is once again, complete with references (in the paper) to prior work demonstrating this point. Indeed, it suggests to me one alternative explanation as to why the Google Desktop Search project was ultimately cancelled – not because it was “no longer necessary” but rather because it wasn’t useful to most computer users.
Another thing that I have observed, repeatedly, when working with students, is there seems to be a natural aversion to searching for answers. Thus, students will post on class forums (such as Piazza, with which I am most familiary) asking questions, even if the question has already been asked and answered. Searching for the answer does not seem to come naturally to such students. Indeed, I often find myself using search engines to find the answer and giving the results back to the original question. I have wondered about this “laziness” in the past and with this new insight I wonder if it is just because people prefer to navigate – and being told where to go is certainly one form of navigation.
Interestingly, like the Graph File System paper I described previously, these authors also argue that we cannot abandon the hierarchical view of files. I am not convinced of this, but I can understand the appeal of starting from it as a basic premise. We have been doing some work recently on a graph visualization model for the file system, more as a prototype, but it is surprisingly functional and encouraging us to look at alternative visualizations of file system data for navigation. In other words, thinking of the problem as a search problem ultimately seems to be the wrong path – yet that is the point of things like semantic file systems, to improve search.
The paper has an extensive review of prior work, much of which I’ve also previous described, though there were a few systems I had not previously seen. Table 1 of the paper has a comparison of features across the various file systems. Thus, the authors distinguish themselves from prior by focusing on providing enhanced functionality, using auxiliary directories, in which they display related content. They focus on:
- Temporal characteristics – they focus on when files are being used, not merely frequency. This is an idea we’ve been exploring as well.
- Geographical location – this is intriguing; identifying where the user was when they accessed a given file.
- File movement – when files are reorganized and moved around.
- File access patterns – they cluster files as related based upon the temporal proximity of their access; another idea we’ve been exploring. I found it insightful they describe this as a “relationship” though they do not explore a broader range of relationships.
- Content similarity – files that are identical or substantially similar can be associated together; this is another technique that we’ve been actively investigating.
- Manual tagging
They describe the file system they implement, which is essentially a layered file system in which they add two virtual directories: NOW and TAGS. They describe the interface for adding this information as well, which I found cumbersome, but it is in keeping with their goal of not deviating from the existing hierarchical interface. They do also permit the creation of custom virtual directories as well, though that is only briefly mentioned in the paper.
One of the problems they highlight, which resonated with me because we’ve been discussing the same problem, is how much information to display to users – in essence, when confronted with too many options, users quickly become overwhelmed.
Their evaluation focuses on the amount of time it took users to locate their files when using their enhanced file systems model and they lay out a case for the fact their system works well for their study group. One limitation of their study group is that it is based upon an experienced computer user group, but this is reflective of their environment.
One interesting comment was that while they used Linux for their evaluation, Windows would have been a better platform because of its broader usage within their organization. I have wondered how much the use of Linux tends to create a bias in an evaluation of this type, since most people are using Windows or Apple computers. Would the results be similar?
The authors do point to their open source implementation of their file system. I have not yet evaluated it, but it is definitely something on my (all too long) list of things to do.
Storage Systems and I/O: Organizing, Storing, and Accessing Data for Scientific Discovery
Storage Systems and I/O: Organizing, Storing, and Accessing Data for Scientific Discovery, Robert Ross, Lee Ward, Philip Carns, Gary Grider, Scott Klasky, Quincey Koziol, Glenn K. Lockwood, Kahthryn Mohror, Bradley Settlemyer, and Matthew Wolf, United States: N. p., 2019. Web. doi:10.2172/1491994.
Google brought this to my attention recently because I’ve set triggers for people citing prior works that I’ve found interesting or useful. There are a number of things that caught my eye with this technical report that apply to my own current research focus. In addition, I received a suggestion that I pick a specific field or discipline and look at how to solve problems in that specific community as an aid to providing focus and motivation for my own work.
This report is a wealth of interesting information and suggestions for areas in which additional work will be useful. The report itself is quite long – 134 pages – so I am really only going to discuss the sections I found useful for framing my own research interests.
The report itself is essentially capturing the discussion at a US DOE workshop. Based upon my reading of the report, it appears that the workshop was structured, with a presentation of information and seeded with topics expected to elicit discussion. Much of the interesting information for me was in Section 4.2 “Metadata, Name Spaces, and Provenance”.
The report starts by defining what they mean by metadata:
Metadata, in this context, refers generally to the information about data. It may include traditional user-visible file system metadata (e.g., file names, permissions, and access times), internal storage system constructs (e.g., data layout information), and extended metadata in support of features such as provenance or user-defined attributes. Metadata access is often characterized by small, latency-bound operations that present a significant challenge for SSIO systems that are optimized for large, bandwidth-intensive transfers. Other challenging aspects of metadata management are the interdependencies among metadata items, consistency requirements of the information about the data, and volume and diversity of metadata workloads.
4.2.1 Metadata (p. 39)
I was particularly interested in seeing their observation that relationships were one of the things that they found missing from existing systems. While somewhat self-serving, this is the observation that has led me into my current research direction. The observation about how meta-data I/O behavior is a mismatch for the bandwidth-intensive nature of accessing the data was a useful insight: there is a real mismatch between the needs of accessing meta-data versus accessing the data itself, particularly for HPC style workloads.
Not explicitly mentioned here is the challenge of constructing a system in which meta-data is an inherent part of the file itself, despite these very different characteristics. The authors do point out other challenges, such as the difficulty in constructing efficient indexing, which suffers from scalability issues:
While most HPC file systems support some notion of extended attributes for files [Braam2002, Welch2008, Weil2006], this type of support is insufficient to capture the desired requirements to establish relationships between distributed datasets, files, and databases; attribute additional complex metadata based on provenance information; and support the mining and analysis of data. Some research systems provide explicit support for searching the file system name space based on attributes [Aviles-Gonzalez2014, Leung2009], but most of these systems rely on effective indexing, which has its own scalability and data-consistency challenges [Chou2011].
4.2.1 Metadata – “State of the Art” (p. 39)
In other words: the performance requirements of accessing meta-data versus data are quite different; the obvious solution is to provide separate storage tiers or services for satisfying these needs. The disadvantage to this is that when we start separating meta-data from data, we create consistency problems and classic solutions to these consistency problems in turn create challenges in scalability. In other words, we are faced with classic distributed systems problems, in which we must trade off consistency versus performance. That is the CAP Theorem in a nutshell.
Another important point in looking at this technical report is that it emphasizes the vast size of datasets in the scientific and HPC communities. These issues of scale are exacerbated in these environments because their needs are extraordinary, with vast data sets, large compute clusters, geographical diversity, and performance demands.
The need for solutions is an important one in this space, as these are definitely pain points. Further, the needs of reproducibility in science are important – the report expressly mentions the DOE policy now requires a data management plan. The emphasis is clearly on reproducibility and data sharing. It seems fairly clear (to me, at least) that having better data sharing can only benefit scientific work.
The seeding discussion for the Metadata section of the report raises some excellent points that, again, help buttress my own arguments and hopefully will be useful in shaping the forward direction:
A number of nontraditional use cases for the metadata management system have emerged as key to DOE missions. These include multiple views of the metadata to support, for example, different views at different levels of the name space hierarchy and different views for different users’ purposes; user-defined metadata; provenance of the metadata; and the ability to define relationships between metadata from different experiments (e.g., to support the provenance use case).
4.2.1 Metadata – “Seeding Workshop Discussion” (p. 40)
As the collection of metadata expands, it is important to ensure that all metadata associated with a dataset remains with the data. Metadata storage at different storage tiers, storage and recovery of metadata from archive, and the transfer of datasets to different storage systems are all important use cases to consider.
The idea of multiple views is important. It is something we’ve been exploring recently as we consider how to look at data just using current information available to us, something I should describe in another post.
So what do I pick out here as being important considerations? Different views, user-defined metadata, provenance, and defining relationships. Their requirement that metadata be associated with the underlying dataset. As I noted previously, this becomes more challenging when you consider that metadata operations have very different interface and performance characteristics than data access.
I have not really been looking at the issues related to metadata management in tiered storage systems, but clearly I have to do so if I want to address the concerns of the HPC community.
The attendees agreed that the added complexity in storage hierarchies presents challenges for locating users’ data. A primary reason is that the community does not yet have efficient mechanisms for representing and querying the metadata of users’ data in a storage hierarchy. Given the bottlenecks that already exist in metadata operations for simple parallel file systems, there is a strong research need to explore how to efficiently support metadata in hierarchical storage systems. A promising direction of research could be to allow users to tag and name their data to facilitate locating the data in the future. The appropriate tagging and naming schemes need investigation and could include information about the data contents to facilitate locating particular datasets, as well as to communicate I/O requirements for the data (e.g., data lifetime or resilience).
Section 4.4.5 Hierarchy and Data Management
How do we efficiently support medatadata in hierarchical storage systems? Implicit in this question is the assumption that there are peculiar challenges for doing this. The report does delve into some of the challenges:
- Hashing can be useful in sharding data for distribution and load balancing, but this does not capture locality – the fact that various files are actually related to one another. We have been considering file clustering; I suspect that cluster sharding might be a useful mechanisms for providing load balancing.
- Metadata generation consists of both automatically collected information (e.g., timestamps, sizes, and name) as well as manually generated information (e.g., tags). The report argues that manual generation is not a particularly effective approach and suggests automatically capturing workflow and provenance information is important. As I was reading this, I was wondering if we might be able to apply inheritance to metadata, in a way that is similar to taint tracking systems.
The report has a short, but useful discussion of namespaces. This includes the traditional POSIX hierarchical name space as well as object oriented name spaces. They point to views as a well-understood approach to the problem from the database community. I would point out the hierarchical approach is one possible view. The report is arguing that their needs would be best met by having multiple views.
The existing work generally is hierarchical and focused on file systems. A number of researchers, however, have argued that such hierarchical namespaces impose inherent limitations on concurrency and usability. Eliminating these limitations with object storage systems or higher-level systems could be the fundamental breakthrough needed to scale namespaces to million-way concurrency and to enable new and more productive interaction modalities.
4.2.2 Namespaces “Seeding Workshop Discussion”
There is a dynamic tension here, between search and navigation. I find myself returning to this issue repeatedly lately and this section reminds me that this is, in fact, an important challenge. Navigation becomes less useful when the namespace becomes large and poorly organized; humans then turn to search. Views become alternative representations of the namespace that humans can use to navigate. They can filter out data that is not useful, which simplifies the task of finding relevant data. We apply views already: we hide “hidden files” or directories beginning with a special character (e.g., “.” in UNIX derived systems). The source code control system git will ignore files (filter them from its view) via a .gitignore file. Thus, we are already applying a primitive, limited form of filtering to create the actual view we show.
This report goes on further. It considers some really interesting issues within this area:
- Storage aware systems for maintaining provenance data.
- The scaling issues inherent in collecting more provenance data; or what do we do when managing the metadata becomes a huge issue itself?
- Cross-system considerations. This doesn’t require HPC data – I have commented more than once that when humans are looking for something, they don’t want to restrict it to the current storage device. Data flows across devices and storage systems; we need to be able to capture these relationships. “[T]here is no formal way to construct, capture, and manage this type of data in an interoperable manner.”
- External meta-data. We need to remember that the context in which the data is collected or consumed is an important aspect of the data itself. Thus, the tools used, the systems, etc. might be factors. I would argue that a storage system can’t reasonably be expected to capture these, but it certainly should be able to store this metadata.
The discussion for this section is equally interesting, because it reflects the thoughts of practitioners and thus their own struggles with the current system:
Attendees mentioned that tracking provenance is a well-explored aspect of many other fields (art history, digital library science, etc.) and that effort should be made to apply the best practices from those fields to our challenges, rather than reinventing them. Attendees extensively discussed the high value of provenance in science reproducibility, error detection and correction in stored data, software fault detection, and I/O performance improvement of current and future systems.
4.2.4 Discussion Themes “Provenance”
Attendees also discussed the need for research into how much provenance information to store, for how long, in what level of detail, and how to ensure that the provenance information was immutable and trustworthy. The value of using provenance beyond strictly validating science data itself was brought up; attendees pointed out that provenance information can be used to train new staff as well as help to retain and propagate institutional knowledge of data gathering processes and procedures.
A generally useful observation: look to how other fields have approached common problems, to see if there are insights from those fields that we can use to address them here. I found the vast reach of the discussion here interesting – the idea that such a system can be used to “… retain and propagate institutional knowledge…”
Finally, I’m going to capture the areas in which the report indicates participants at the workshop reached consensus. I’ll paraphrase, rather than quote:
- Scalable metadata storage – a key point for me here was decoupling meta-data from the data itself. That, despite the seeding suggestion that we keep meta-data associated with the file.
- Improve namespace query and display capabilities – make them dynamic, programmable, and extensible.
- Better provenance information – the emphasis was on reproducibility of results, but they wanted to ensure that this could embed domain specific features, so that such systems can be useful beyond just reproducibility.
I’ve really only touched on a small part of this report’s total content; there is quite a bit of other, useful, insight within the report. I will be mulling over these issues
Visualization
Last post I discussed relationships. But relationships really are not enough. Another key to this puzzle is visualization. In other words, how do we present the information to users so that it is useful.
But first, let me step back and point to a larger problem: information overload. If we present users with a list of 100,000 options, they won’t be able to necessarily find what they’re actually seeking. For example, one of the challenges of using an Internet search engine is that they can return millions of “answers” to my basic query. In fact, I just typed in “What is the meaning of life?” to a search engine and it responded back with 1.1 billion possible answers, all in under a second. They are a marvel at backing up the dump truck and offering to inundate me with answers. When is the last time you went through even a handful of these, let alone a large percentage of them?
I suspect that if I ask the HCI folks they will be able to tell me what works for ordinary mortals, but I assure you that computers are capable of returning more information than one can possibly ever process – many years ago I received a bug report about a file system directory that contained over 700,000 files and did not display with the file system kit I’d constructed and the company sold. I’d known I made decisions about resources when I did it; but we lifted that restriction and supported much larger directories than that. I’m quite sure that no mere human would look at such a directory in any meaningful sense. Maybe it was a collection of log files, in which case the names likely embed semantic information about the files themselves. Indeed, in Burrito the authors noted that scientists often embed the schema of their data within the file names. I know that I have done the same and when I’m looking for specific data I am often scripting code to sift through the pile of data to find the subset that is useful to me. The point remains the same: huge listings of files within a directory don’t work for humans.
One potential area for considering navigation is faceted search, a technique for making vast quantities of data searchable. Indeed, this fits well with the graph file system idea because what we expect to find in our graphs are clusters of related files. The graph is likely to be a sparse graph because most files are unlikely to share common features with one another. Thus, it suggests that at least one model for this data visualization problem is going to be rolling up data into these clusters, with an iterative approach to breaking it down and displaying it further. Of course, we might be able to do that within the confines of the existing hierarchical structure, which would be great for retrofitting this into the vast array of existing applications; still, my hope is that we can also provide novel new variations (or rather someone more clever than I am at these things will do so). The challenge is to build something that enables this means I need to have some level of understanding as to what kinds of information are needed to do so.
For example, I was wondering about first order approximations – those that mimic an existing hierarchical file system. Such a file system would present one or more views of the data. Maybe we have a Time view, and the time view then shows you all the files in time order. But if you have 1,000,000 files, that is going to be a mighty big list. One option might be to divide it up into ever smaller chunks of time. Eventually, we’d get to a point where you could see a few dozen files in some sort of time order. Of course, strict sharding of time might not make sense either: why should two files that are separated in time by small intervals end up in separate locations.
One possible option would be to consider a time slider that controls the view. This is something that we can find in other temporal data tools. Thus, creating a time slider might help make visualization easier to perform – in some ways this is similar to the Windows timeline feature that has recently been added into Windows 10. I suspect the file system doesn’t do much to facilitate this. From my own use of this feature it has some interesting limitations, not the least of which is that if you want full functionality they want to export data “to the cloud” for further analysis. That sounds like whatever they are doing is data intensive, which in turn suggests to me that the file system is not doing anything to facilitate this. If the answer really is “sorry, but you have to spend lots of computational cycles to mine this data” then my research is unlikely to be fruitful. I push forward because I don’t think this really is the case – the database community has done marvellous things that permit a vast treasure of relationships be mined.
If we combine this with a faceted search area, I can envision a filtered timeline model – in essence, this seems to be what the new Windows feature is doing, albeit by filtering the things they have deemed to be of interest. I suspect they will extend this capability over time, but a time ordered view is just one of the possible relationships I consider to be important for consideration. I don’t know what the relationships will be, but I do think that having a pre-defined list of those relationships will be self-limiting. Perhaps it will be enough. I proceed on the basis that I expect it will not be enough.
One possible visualization I’ve been considering – and part of the motivation for me deciding I need to start building a file system – is that this sort of namespace might be achievable on our existing hierarchical model if I just add “an extra level of indirection”. Suppose we construct a file system that, instead of having the existing file name has a directory of the same name. In turn, that directory then contains relationships associated with other objects. One of those other objects could be the actual file (so we can still access it), but we could also have a “temporal” directory that would then display a list of files that were created, modified, or accessed around the same timeframe. We could store information about what web sites were visited around times that the file was in use. We could keep track of the music that was playing around the same time. We could point to files that were similar to that specific file. Such a visualization could be easily achieved and compatible with existing tools. Rather than being an endpoint, this is more an intermediate staging area – a way of mocking up the concepts and ideas, and to see what works for people and what does not work.
Thus the desire for a file system. I think we can construct a static namespace by using an existing file system and symbolic links (so you can eventually get back to the real files) by mining existing data sources. But eventually, we are going to want dynamic support here. We can stub that out with FUSE (for example) but in my experience (and based upon the literature) FUSE is slow for meta-data operations, which is really all I will be doing. Building file systems is hard work, but it is something I’ve been doing for years, so that aspect doesn’t really scare me. Visualization on the other hand is an area in which I don’t have a lot of experience.
I’m certainly open to ideas…
Recent Comments