BEGIN:VCALENDAR VERSION:2.0 X-WR-CALNAME:EventsCalendar PRODID:-//hacksw/handcal//NONSGML v1.0//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York LAST-MODIFIED:20240422T053451Z TZURL:https://www.tzurl.org/zoneinfo-outlook/America/New_York X-LIC-LOCATION:America/New_York BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT CATEGORIES:College of Engineering,Graduate Studies,Lectures and Seminars,Th esis/Dissertations DESCRIPTION:Faculty Supervisor: Dr. Gokhan Kul, Computer & Information Scie nce Committee Members: Dr. Adnan El-Nasan, Computer & Information ScienceD r. John Rahme, Computer & Information Science Abstract: Memory allocators perform a fundamental task in ensuring dynamic memory is allocated to pro cesses in a way that is efficient in both time and space domains. They do this by reducing the number of kernel calls performed, maintaining high sp atial and cache locality, and keeping the number of pages in use to a mini mum. However, there is a fundamental constraint towards the efficiency of an allocator. External fragmentation is an unavoidable consequence of dyna mic allocation, yet there is little consensus on how precisely to measure fragmentation; There is vanishingly little analysis on how different fragm entation metrics perform or when each should be used. A better understandi ng of fragmentation and how to measure it is vital in advancing allocator design and designing programs that respect allocators’ assumptions.  Th ere is a large myriad of work introducing various metrics for measuring fr agmentation or describing allocators that reduce fragmentation in some nov el way under particular workloads. However, there is a gap in modern work determining the efficacy of the numerous proposed fragmentation metrics, w hen they should be used, and how to compare them between each other. Many papers discussing fragmentation produced under some workload or allocator don’t describe which metric is being used. This disjoint approach to mea suring fragmentation leads to confusion and imprecision. Much of the work on fragmentation is old, dating back 50 years. Newer work tends to be focu sed on allocation on either large clusters or databases, or in optical net work allocation rather than memory. Additionally, there has been a focus o n analyzing synthetic traces (and how precisely to generate synthetic trac es), rather than using traces from real workloads on real machines. This w ork attempts to unify the understanding of fragmentation through comparati ve analysis on real workloads. This work introduces a pipeline for benchma rking external fragmentation metrics using real memory traces. First is a method of collecting memory traces of multi-threaded and multi-process sys tems, tracking every call to the allocation library. Next a method of comp uting the fragmentation of these traces at variable timesteps through the parent process’ lifetime. And finally, a method of plotting metrics over time against each other and against the overall memory usage of that proc ess. 6 metrics are compared in this work, including two novel metrics; AEF M and NAEFM, and one metric that has escaped academic review; ESP. The ESP metric performs extremely well, holding high average correlation with oth er metrics, while maintaining much more stability on shorter applications. For longer running applications, many metrics quickly lose specificity, r eaching their maximum value quite quickly and staying there. NAEFM, SSFM, and EBFM maintain usefulness across the lifetime of longer running program s.  This work has applications in a wide range of areas. It is naturally of interest to kernel or allocator developers. Future work which takes adv antage of findings from this comparative analysis may produce more efficie nt allocators on small and personal computers. A better understanding of m emory degradation over time will result in more efficient applications, re ducing the overall resource burden of everyday computing. Additionally, th ere may be applications in adjacent fields as found with optical network a llocation.   For further information please contact Dr. Gokhan Kul at gku l@umassd.edu.    \nEvent page: https://www.umassd.edu/events/cms/benchm arking-of-memory-fragmentation-metrics-on-real-allocation-traces.php\nEven t link: https://teams.microsoft.com/meet/28305274898079?p=Hq5YC1bJl1SW8ioW En X-ALT-DESC;FMTTYPE=text/html:

Faculty Supervisor:
Dr. G okhan Kul\, Computer & Information Science

\n

Committee Members:
Dr. Adnan El-Nasan\, Computer & Information Science
Dr. John Rahme\ , Computer & Information Science

\n

Abstract: Memor y allocators perform a fundamental task in ensuring dynamic memory is allo cated to processes in a way that is efficient in both time and space domai ns. They do this by reducing the number of kernel calls performed\, mainta ining high spatial and cache locality\, and keeping the number of pages in use to a minimum. However\, there is a fundamental constraint towards the efficiency of an allocator. External fragmentation is an unavoidable cons equence of dynamic allocation\, yet there is little consensus on how preci sely to measure fragmentation\; There is vanishingly little analysis on ho w different fragmentation metrics perform or when each should be used. A b etter understanding of fragmentation and how to measure it is vital in adv ancing allocator design and designing programs that respect allocators’ assumptions.  There is a large myriad of work introducing various metrics for measuring fragmentation or describing allocators that reduce fragment ation in some novel way under particular workloads. However\, there is a g ap in modern work determining the efficacy of the numerous proposed fragme ntation metrics\, when they should be used\, and how to compare them betwe en each other. Many papers discussing fragmentation produced under some wo rkload or allocator don’t describe which metric is being used. This disj oint approach to measuring fragmentation leads to confusion and imprecisio n. Much of the work on fragmentation is old\, dating back 50 years. Newer work tends to be focused on allocation on either large clusters or databas es\, or in optical network allocation rather than memory. Additionally\, t here has been a focus on analyzing synthetic traces (and how precisely to generate synthetic traces)\, rather than using traces from real workloads on real machines. This work attempts to unify the understanding of fragmen tation through comparative analysis on real workloads. This work introduce s a pipeline for benchmarking external fragmentation metrics using real me mory traces. First is a method of collecting memory traces of multi-thread ed and multi-process systems\, tracking every call to the allocation libra ry. Next a method of computing the fragmentation of these traces at variab le timesteps through the parent process’ lifetime. And finally\, a metho d of plotting metrics over time against each other and against the overall memory usage of that process. 6 metrics are compared in this work\, inclu ding two novel metrics\; AEFM and NAEFM\, and one metric that has escaped academic review\; ESP. The ESP metric performs extremely well\, holding hi gh average correlation with other metrics\, while maintaining much more st ability on shorter applications. For longer running applications\, many me trics quickly lose specificity\, reaching their maximum value quite quickl y and staying there. NAEFM\, SSFM\, and EBFM maintain usefulness across th e lifetime of longer running programs.  This work has applications in a w ide range of areas. It is naturally of interest to kernel or allocator dev elopers. Future work which takes advantage of findings from this comparati ve analysis may produce more efficient allocators on small and personal co mputers. A better understanding of memory degradation over time will resul t in more efficient applications\, reducing the overall resource burden of everyday computing. Additionally\, there may be applications in adjacent fields as found with optical network allocation.  

\n

For further in formation please contact Dr. Gokhan Kul at gkul@umassd.edu.  

\n

 

Event page: https://www.umassd.edu/events/cms/benchmarking- of-memory-fragmentation-metrics-on-real-allocation-traces.php
Event link: https://teams.microsoft.com/meet/28305274898079?p=Hq5YC1bJl1SW 8ioWEn

DTSTAMP:20260421T152010 DTSTART;TZID=America/New_York:20260511T100000 DTEND;TZID=America/New_York:20260511T110000 LOCATION:Dion 311 and Teams (https://teams.microsoft.com/meet/2830527489807 9?p=Hq5YC1bJl1SW8ioWEn) SUMMARY;LANGUAGE=en-us:Benchmarking of Memory Fragmentation Metrics on Real Allocation Traces UID:d6a21cbca7ae90810049e48d7c168313@www.umassd.edu END:VEVENT END:VCALENDAR