Decoupling GPU Programming Models from Resource Management for Enhanced Programming Ease, Portability, and Performance

05/02/2018 ∙ by Nandita Vijaykumar, et al. ∙ 0

The application resource specification--a static specification of several parameters such as the number of threads and the scratchpad memory usage per thread block--forms a critical component of modern GPU programming models. This specification determines the parallelism, and hence performance, of the application during execution because the corresponding on-chip hardware resources are allocated and managed based on this specification. This tight-coupling between the software-provided resource specification and resource management in hardware leads to significant challenges in programming ease, portability, and performance. Zorua is a new resource virtualization framework, that decouples the programmer-specified resource usage of a GPU application from the actual allocation in the on-chip hardware resources. Zorua enables this decoupling by virtualizing each resource transparently to the programmer. We demonstrate that by providing the illusion of more resources than physically available via controlled and coordinated virtualization, Zorua offers several important benefits: (i) Programming Ease. Zorua eases the burden on the programmer to provide code that is tuned to efficiently utilize the physically available on-chip resources. (ii) Portability. Zorua alleviates the necessity of re-tuning an application's resource usage when porting the application across GPU generations. (iii) Performance. By dynamically allocating resources and carefully oversubscribing them when necessary, Zorua improves or retains the performance of applications that are already highly tuned to best utilize the resources.



There are no comments yet.


page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

This document provides instructions for submitting papers to the 49th International Symposium on microarchitecture (MICRO), 2016. In an effort to respect the efforts of reviewers and in the interest of fairness to all prospective authors, we request that all submissions to MICRO 2016 follow the formatting and submission rules detailed below. Submissions that violate these instructions may not be reviewed, at the discretion of the program chair, in order to maintain a review process that is fair to all potential authors.

An example file (formatted using the MICRO’16 submission format) that contains the formatting guidelines can be downloaded from here: Sample PDF. The content of this document mirrors that of the submission instructions that appear on this website, where the paper submission site will be linked online shortly.

All questions regarding paper formatting and submission should be directed to the program chair.

1.1 Format Highlights

Note that there are some changes from last year.

  • Paper must be submitted in printable PDF format.

  • Text must be in a minimum 10pt (not 9pt) font.

  • Papers must be at most 11 pages, not including references.

  • No page limit for references.

  • Each reference must specify all authors (no et al.).

  • Authors may optionally suggest reviewers.

  • Authors of all accepted papers will be required to give a lightning presentation (about 90s) and a poster in addition to the regular conference talk.

1.2 Paper Evaluation Objectives

The committee will make every effort to judge each submitted paper on its own merits. There will be no target acceptance rate. We expect to accept a wide range of papers with appropriate expectations for evaluation — while papers that build on significant past work with strong evaluations are valuable, papers that open new areas with less rigorous evaluation are equally welcome and especially encouraged. Given the wide range of topics covered by MICRO, every effort will be made to find expert reviewers, including providing the ability for authors’ to suggest additional reviewers.

2 Paper Preparation Instructions

2.1 Paper Formatting

Papers must be submitted in printable PDF format and should contain a maximum of 11 pages of single-spaced two-column text, not including references. You may include any number of pages for references, but see below for more instructions. If you are using LaTeX [lamport94] to typeset your paper, then we suggest that you use the template here: LaTeX Template. This document was prepared with that template. If you use a different software package to typeset your paper, then please adhere to the guidelines given in Table 1.

Field Value
File format PDF
Page limit 11 pages, not including
Paper size US Letter 8.5in 11in
Top margin 1in
Bottom margin 1in
Left margin 0.75in
Right margin 0.75in
Body 2-column, single-spaced
Space between columns 0.25in
Body font 10pt
Abstract font 10pt, italicized
Section heading font 12pt, bold
Subsection heading font 10pt, bold
Caption font 9pt (minimum), bold
References 8pt, no page limit, list
all authors’ names
Table 1: Formatting guidelines for submission.

Please ensure that you include page numbers with your submission. This makes it easier for the reviewers to refer to different parts of your paper when they provide comments.

Please ensure that your submission has a banner at the top of the title page, similar to this one, which contains the submission number and the notice of confidentiality. If using the template, just replace XXX with your submission number.

2.2 Content

Author List. Reviewing will be double blind; therefore, please do not include any author names on any submitted documents except in the space provided on the submission form. You must also ensure that the metadata included in the PDF does not give away the authors. If you are improving upon your prior work, refer to your prior work in the third person and include a full citation for the work in the bibliography. For example, if you are building on your own prior work in the papers [nicepaper1, nicepaper2, nicepaper3], you would say something like: ”While the authors of [nicepaper1, nicepaper2, nicepaper3] did X, Y, and Z, this paper additionally does W, and is therefore much better.” Do NOT omit or anonymize references for blind review. There is one exception to this for your own prior work that appeared in IEEE CAL, workshops without archived proceedings, etc.  as discussed later in this document.

Figures and Tables. Ensure that the figures and tables are legible. Please also ensure that you refer to your figures in the main text. Many reviewers print the papers in gray-scale. Therefore, if you use colors for your figures, ensure that the different colors are highly distinguishable in gray-scale.

References. There is no length limit for references. Each reference must explicitly list all authors of the paper. Papers not meeting this requirement will be rejected. Authors of NSF proposals should be familiar with this requirement. Knowing all authors of related work will help find the best reviewers. Since there is no length limit for the number of pages used for references, there is no need to save space here.

3 Paper Submission Instructions

3.1 Guidelines for Determining Authorship

IEEE guidelines dictate that authorship should be based on a substantial intellectual contribution. It is assumed that all authors have had a significant role in the creation of an article that bears their names. In particular, the authorship credit must be reserved only for individuals who have met each of the following conditions:

  1. Made a significant intellectual contribution to the theoretical development, system or experimental design, prototype development, and/or the analysis and interpretation of data associated with the work contained in the article;

  2. Contributed to drafting the article or reviewing and/or revising it for intellectual content; and

  3. Approved the final version of the article as accepted for publication, including references.

A detailed description of the IEEE authorship guidelines and responsibilities is available here. Per these guidelines, it is not acceptable to award honorary authorship or gift authorship. Please keep these guidelines in mind while determining the author list of your paper.

3.2 Declaring Authors

Declare all the authors of the paper upfront. Addition/removal of authors once the paper is accepted will have to be approved by the program chair, since it potentially undermines the goal of eliminating conflicts for reviewer assignment.

3.3 Areas and Topics

Authors should indicate these areas on the submission form as well as specific topics covered by the paper for optimal reviewer match. If you are unsure whether your paper falls within the scope of MICRO, please check with the program chair – MICRO is a broad, multidisciplinary conference and encourages new topics.

3.4 Declaring Conflicts of Interest

Authors must register all their conflicts on the paper submission site. Conflicts are needed to ensure appropriate assignment of reviewers. If a paper is found to have an undeclared conflict that causes a problem OR if a paper is found to declare false conflicts in order to abuse or “game” the review system, the paper may be rejected.

We use the NSF conflict of interest guidelines for determining the conflict period for MICRO’15. Please declare a conflict of interest (COI) with the following people for any author of your paper:

  1. Your Ph.D. advisor(s), post-doctoral advisor(s), Ph.D. students, and post-doctoral advisees, forever.

  2. Family relations by blood or marriage, or their equivalent, forever (if they might be potential reviewers).

  3. People with whom you have collaborated in the last FOUR years, including

    • co-authors of accepted/rejected/pending papers.

    • co-PIs on accepted/rejected/pending grant proposals.

    • funders (decision-makers) of your research grants, and researchers whom you fund.

  4. People (including students) who shared your primary institution(s) in the last FOUR years.

  5. Other relationships, such as close personal friendship, that you think might tend to affect your judgment or be seen as doing so by a reasonable person familiar with the relationship.

“Service” collaborations such as co-authoring a report for a professional organization, serving on a program committee, or co-presenting tutorials, do not themselves create a conflict of interest. Co-authoring a paper that is a compendium of various projects with no true collaboration among the projects does not constitute a conflict among the authors of the different projects.

On the other hand, there may be others not covered by the above with whom you believe a COI exists, for example, an ongoing collaboration which has not yet resulted in the creation of a paper or proposal. Please report such COIs; however, you may be asked to justify them. Please be reasonable. For example, you cannot declare a COI with a reviewer just because that reviewer works on topics similar to or related to those in your paper. The PC Chair may contact co-authors to explain a COI whose origin is unclear.

We hope to draw most reviewers from the PC and the ERC, but others from the community may also write reviews. Please declare all your conflicts (not just restricted to the PC and ERC). When in doubt, contact the program chair.

3.5 Optional Reviewer Suggestions

Authors may optionally mark (non-conflicted) PC and ERC members that they believe could provide expert reviews for their submission. If authors believe there is insufficient expertise on the PC and ERC for the topic of their paper, they may suggest alternate reviewers. The program chair will use the authors’ input at his discretion. We provide this opportunity for input mostly for papers on non-traditional and emerging topics.

3.6 Concurrent Submissions and Workshops

By submitting a manuscript to MICRO’15, the authors guarantee that the manuscript has not been previously published or accepted for publication in a substantially similar form in any conference, journal, or the archived proceedings of a workshop (e.g., in the ACM digital library) – see exceptions below. The authors also guarantee that no paper that contains significant overlap with the contributions of the submitted paper will be under review for any other conference or journal or an archived proceedings of a workshop during the MICRO’15 review period. Violation of any of these conditions will lead to rejection.

The only exceptions to the above rules are for the authors’ own papers in (1) workshops without archived proceedings such as in the ACM digital library (or where the authors chose not to have their paper appear in the archived proceedings), or (2) venues such as IEEE CAL where there is an explicit policy that such publication does not preclude longer conference submissions. In all such cases, the submitted manuscript may ignore the above work to preserve author anonymity. This information must, however, be provided on the submission form – the PC chair will make this information available to reviewers if it becomes necessary to ensure a fair review. As always, if you are in doubt, it is best to contact the program chair.

Finally, we also note that the ACM Plagiarism Policy ( covers a range of ethical issues concerning the misrepresentation of other works or one’s own work.

4 Acknowledgements

This document is derived from previous conferences, in particular MICRO 2013 and ASPLOS 2015. We thank Christos Kozyrakis and Sandhya Dwarkadas for their inputs.