2026-09-19T03:29:29Zhttps://dspace.mit.edu/server/oai/requestoai:dspace.mit.edu:1721.1/1512732023-08-01T03:08:28Zcom_1721.1_7582com_1721.1_7581col_1721.1_131023
Schardl, Tao B.
Govedic, Luka
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
2023-07-31T19:27:42Z
2023-07-31T19:27:42Z
2023-06
2023-06-06T16:34:32.151Z
https://hdl.handle.net/1721.1/151273
For good performance, parallel loop scheduling must achieve low scheduling overheads and multidimensional locality in nested loops. This thesis explores both challenges and contributes an extension to randomized work-stealing for first-class loop support that reduces scheduling overheads.
Randomized work-stealing schedulers traditionally execute parallel-for loops using parallel divide-and-conquer recursion, which is theoretically efficient and scalable but can incur substantial overheads in practice. This thesis extends randomized work-stealing with a custom work-stealing protocol called on-the-fly loop splitting. I introduce loop frames to make work stealing on parallel-for loops more efficient and flexible.
Loop frames make two key changes to work stealing for parallel-for loops. First, loop frames extend work stealing by directly encoding information about intervals of loop iterations in the runtime. Loop frames add first-class support to work stealing for parallel-for loops that composes with classical randomized work stealing. Second, loop frames allow intervals of loop iterations to be split on-the-fly, such that worker threads attempt to steal half of the unexecuted loop iterations rather than a deterministically constructed partition of loop iterations. On-the-fly loop splitting allows for more flexible dynamic load balancing of loop iterations while keeping the work overheads low and maintaining the theoretical efficiency of divide-and-conquer.
I evaluate loop frames in practice by implementing loop frames in the OpenCilk runtime system. In particular, loop frames augment the THE protocol from Cilk to coordinate updates to loop frames. I observe that loop frames and on-the-fly loop splitting incur substantially less overhead than the divide-and-conquer algorithm without sacrificing parallel scalability.
Finally, I study the impacts of increased locality in more than one dimension in nested loop applications. Results show that both cache-aware and cache-oblivious reordering of nested loop iterations can result in performance benefits up to a factor of 1.7×.
M.Eng.
Massachusetts Institute of Technology
In Copyright - Educational Use Permitted
Copyright retained by author(s)
https://rightsstatements.org/page/InC-EDU/1.0/
Improving the Performance of Parallel Loops in OpenCilk
Thesis
application/pdf
Master
Master of Engineering in Electrical Engineering and Computer Science
Publication
unknown
unknown
<Publication xmlns="https://www.openaire.eu/cerif-profile/1.1/" id="dba23bfb-7699-4032-b133-81da23bc0d45">
<Type xmlns="https://www.openaire.eu/cerif-profile/vocab/COAR_Publication_Types">http://purl.org/coar/resource_type/c_1843</Type>
<Title>Improving the Performance of Parallel Loops in OpenCilk</Title>
<PublishedIn>
<Publication>
</Publication>
</PublishedIn>
<PublicationDate>2023-06</PublicationDate>
<Authors>
<Author>
<DisplayName>Govedic, Luka</DisplayName>
<Affiliation>
<OrgUnit>
</OrgUnit>
</Affiliation>
</Author>
</Authors>
<Editors>
</Editors>
<Publishers>
<Publisher>
<DisplayName>Massachusetts Institute of Technology</DisplayName>
<OrgUnit />
</Publisher>
</Publishers>
<License>https://rightsstatements.org/page/InC-EDU/1.0/</License>
<Abstract>For good performance, parallel loop scheduling must achieve low scheduling overheads and multidimensional locality in nested loops. This thesis explores both challenges and contributes an extension to randomized work-stealing for first-class loop support that reduces scheduling overheads.
Randomized work-stealing schedulers traditionally execute parallel-for loops using parallel divide-and-conquer recursion, which is theoretically efficient and scalable but can incur substantial overheads in practice. This thesis extends randomized work-stealing with a custom work-stealing protocol called on-the-fly loop splitting. I introduce loop frames to make work stealing on parallel-for loops more efficient and flexible.
Loop frames make two key changes to work stealing for parallel-for loops. First, loop frames extend work stealing by directly encoding information about intervals of loop iterations in the runtime. Loop frames add first-class support to work stealing for parallel-for loops that composes with classical randomized work stealing. Second, loop frames allow intervals of loop iterations to be split on-the-fly, such that worker threads attempt to steal half of the unexecuted loop iterations rather than a deterministically constructed partition of loop iterations. On-the-fly loop splitting allows for more flexible dynamic load balancing of loop iterations while keeping the work overheads low and maintaining the theoretical efficiency of divide-and-conquer.
I evaluate loop frames in practice by implementing loop frames in the OpenCilk runtime system. In particular, loop frames augment the THE protocol from Cilk to coordinate updates to loop frames. I observe that loop frames and on-the-fly loop splitting incur substantially less overhead than the divide-and-conquer algorithm without sacrificing parallel scalability.
Finally, I study the impacts of increased locality in more than one dimension in nested loop applications. Results show that both cache-aware and cache-oblivious reordering of nested loop iterations can result in performance benefits up to a factor of 1.7×.</Abstract>
<Access xmlns="http://purl.org/coar/access_right"
>
</Access>
</Publication>