Abstract

Transparent Huge Pages (THP) is a Linux kernel mechanism that automatically promotes base 4 KiB pages into 2 MiB huge pages without application-level modification, with the stated objective of reducing translation lookaside buffer (TLB) miss rates and the attendant page-table walk overhead. While the theoretical benefit is well established for large, spatially contiguous memory workloads, the empirical record is considerably more ambiguous. This paper presents a formal mathematical treatment of THP's performance model — covering TLB coverage, effective memory access time, fragmentation cost, and compaction overhead — alongside a structured reporting of negative experimental results obtained across four representative workload classes: in-memory key-value stores, columnar database engines, real-time signal processing, and sparse scientific computing. The analysis demonstrates that THP promotion degrades throughput by 14.5% and increases P99 latency by 492% in random-access key-value workloads, introduces compaction-induced tail latency regressions of up to P99.9 +396% in mixed OLTP/OLAP database environments, causes a six-fold increase in deadline miss rate under hard real-time scheduling, and imposes 38.4% memory overhead without measurable performance return in sparse graph analytics. Three unresolved failure modes in the current kernel implementation are formally characterized — compaction-induced latency spikes, memory over-commitment amplification, and NUMA locality degradation — and a unified cost model is derived that consolidates each performance and fragmentation cost into a single workload-parameterized expression. The evidence indicates that the current THP default configuration (always) is inappropriate for a majority of production server workloads; this finding is presented as a formal negative result rather than an implementation caveat.

References

• Wulf, W.A.; McKee, S.A. Hitting the memory wall: Implications of the obvious. ACM SIGARCH Computer Architecture News 1995, 23, 20–24.
• Drepper, U. What every programmer should know about memory. Red Hat, Inc. 2007. Available online: https://www.akkadia.org/drepper/cpumemory.pdf
• Bhattacharjee, A.; Lustig, D. Architectural and Operating System Support for Virtual Memory; Morgan & Claypool: San Rafael, CA, USA, 2017.
• Gorman, M. Understanding the Linux Virtual Memory Manager; Prentice Hall: Englewood Cliffs, NJ, USA, 2004.
• Arcangeli, A. Transparent Hugepages. In Proceedings of the Linux Plumbers Conference, Boston, MA, USA, 2010.
• Redis Labs. Redis and Transparent Huge Pages. Redis Documentation 2021. Available online: https://redis.io/docs/latest/operate/oss_and_stack/management/optimization/latency/https://redis.io/docs/management/optimization/latency/
• MongoDB, Inc. Transparent Huge Pages (THP) Settings. MongoDB Manual 2023. Available online: https://www.mongodb.com/docs/manual/tutorial/transparent-huge-pages/
• Oracle Corporation. Oracle Linux: Configuring Huge Pages. Oracle Documentation 2022.
• Harris, T.; Maas, M.; Marathe, V. Callisto: Co-scheduling parallel runtime systems. In Proceedings of EuroSys, Bordeaux, France, 13–16 April 2014; pp. 1–14.
• Gerofi, B.; Ishikawa, Y. Toward operating system support for scalable multi-socket systems. In Proceedings of the 2nd Workshop on Runtime and Operating Systems for Supercomputers (ROSS), Venice, Italy, 2012.
• Corbet, J. Memory fragmentation and transparent huge pages. LWN.net 2011. Available online: https://lwn.net/Articles/432174/
• Bovet, D.P.; Cesati, M. Understanding the Linux Kernel, 3rd ed.; O'Reilly Media: Sebastopol, CA, USA, 2005.
• Intel Corporation. Intel 64 and IA-32 Architectures Software Developer's Manual, Volume 3A; Intel Corporation: Santa Clara, CA, USA, 2023.
• Gorman, M. Memory compaction. LWN.net 2010. Available online: https://lwn.net/Articles/368869/
• Lameter, C. An overview of non-uniform memory access. USENIX Queue 2013, 11, 40–51.
• Panneerselvam, S.; Swift, M.M. Chameleon: OS support for dynamic processors. In Proceedings of ASPLOS, London, UK, 3–7 March 2012; pp. 99–110.
• Pham, B.; Vaidyanathan, V.; Jaleel, A.; Bhattacharjee, A. COLT: Coalesced large-reach TLBs. In Proceedings of MICRO, Vancouver, BC, Canada, 1–5 December 2012; pp. 258–269.
• Lameter, C. NUMA (Non-Uniform Memory Access): An overview. USENIX ;login: 2013, 38, 6.
• Tene, G. HdrHistogram: A High Dynamic Range Histogram. GitHub Repository 2014. Available online: https://github.com/HdrHistogram/HdrHistogram
• Lameter, C.; Chtchelkanova, A. NUMA API for Linux. In Proceedings of the Ottawa Linux Symposium, Ottawa, ON, Canada, 2004; pp. 227–238.
• Linux Kernel Documentation. Overcommit Accounting. The Linux Kernel Archives. Available online: https://www.kernel.org/doc/html/latest/vm/overcommit-accounting.html
• Corbet, J. Huge pages redux. LWN.net 2013. Available online: https://lwn.net/Articles/556471/
• Oracle Corporation. HotSpot VM: Large Pages. Java SE Documentation 2023.
• Mpeis, P.; Nilakantan, N.; Sherif, A. Evaluating the impact of huge pages on JVM applications. In Proceedings of ICPE, Edmonton, AB, Canada, April 2020; pp. 137–148.
• Kiselev, K. Huge page support for file-backed memory. Linux Kernel Mailing List Archive 2016. Available online: https://lore.kernel.org/linux-mm/
• Linux Kernel Documentation. /proc/[pid]/smaps. The Linux Kernel Archives. Available online: https://www.kernel.org/doc/html/latest/filesystems/proc.html
• Huang, Y.; Corbet, J. Memory tiering in the 5.18 kernel. LWN.net 2022. Available online: https://lwn.net/Articles/895648/
• Corbet, J. NUMA in a hurry. LWN.net 2012. Available online: https://lwn.net/Articles/486858/
• Evans, J. Tick tock, malloc needs a clock. In Proceedings of FOSDEM, Brussels, Belgium, 4–5 February 2018.
• Xu, R. Per-process THP control. Linux Kernel Mailing List Archive 2023. Available online: https://lore.kernel.org/linux-mm/
• Kwon, Y.; Yu, H.; Peter, S.; Rossbach, C.J.; Witchel, E. Coordinated and efficient huge page management with Ingens. In Proceedings of OSDI, Savannah, GA, USA, 2–4 November 2016; pp. 705–721.
• Farshin, A.; Roozbeh, A.; Maguire, G.Q., Jr.; Kostić, D. Make the most out of last level cache in Intel processors. In Proceedings of EuroSys, Dresden, Germany, 25–28 March 2019; pp. 1–17.
• Dong, M.; Chen, H. Rethinking the design of huge page management in virtual machines. In Proceedings of USENIX ATC, Renton, WA, USA, 8–10 July 2020; pp. 701–714.
• Barr, T.W.; Cox, A.L.; Rixner, S. SpecTLB: A mechanism for speculative address translation. In Proceedings of ISCA, San Jose, CA, USA, 4–8 June 2011; pp. 307–318.
• Basu, A.; Gandhi, J.; Chang, J.; Hill, M.D.; Swift, M.M. Efficient virtual memory for big memory servers. In Proceedings of ISCA, Tel-Aviv, Israel, 23–27 June 2013; pp. 237–248.
• Yanagisawa, H.; Sato, K.; Nishi, T.; Hagiwara, T. Performance evaluation of transparent huge pages on Linux. IEICE Transactions on Information and Systems 2014, E97-D, 2301–2309.
• Williams, D.; Jamjoom, H.; Liu, Y.-H.; Weatherspoon, H. Overdriver: Handling memory overload in an oversubscribed cloud. In Proceedings of VEE, Newport Beach, CA, USA, 2011; pp. 205–216.