Hoard (memory allocator)

Hoard (memory allocator)
Name	Hoard
Type	Memory allocator
Developer	Emery Berger and team
First release	2000
Programming language	C, C++
License	BSD-like

Hoard (memory allocator) is a scalable memory allocator designed to reduce contention and fragmentation in multithreaded programs. It was developed to address allocator-induced bottlenecks observed in server-class workloads and desktop applications, balancing per-thread performance with global memory efficiency. Hoard influenced allocator research and production implementations by emphasizing cache-locality, low-locking, and heap-coordination strategies.

History

Hoard originated from research led by Emery Berger at the University of Massachusetts Amherst in collaboration with colleagues and students, motivated by concurrency issues identified in experiments with workloads from projects at Microsoft Research, Sun Microsystems, and academic benchmarks such as SPEC. Early publications presented Hoard at venues like USENIX, ACM OOPSLA, and SIGPLAN workshops, drawing citations from later work at institutions including Stanford University, Massachusetts Institute of Technology, Carnegie Mellon University, and University of California, Berkeley. Funding and interest came from agencies and labs such as the National Science Foundation and collaborations with industrial groups at Intel Corporation and IBM Research. Over time, Hoard informed allocators in systems produced by companies like Google, Facebook, and Oracle Corporation, while being incorporated into comparative studies alongside allocators developed at Red Hat and experimental systems from MIT CSAIL.

Design and Architecture

Hoard uses a layered design that separates per-thread allocation fast paths from a centralized heap coordination mechanism. The architecture includes per-CPU or per-thread heaps influenced by ideas from Andrew Birrell’s work on per-processor resource management and concepts explored at Bell Labs research on concurrent data structures. It employs locking and lock-free primitives studied in literature from Herlihy and Shavit, combined with fragmentation bounds related to analyses by researchers at Princeton University and Cornell University. The allocator organizes memory into superblocks and size classes similar to strategies used by teams at Mozilla Foundation and Facebook for their allocators, while maintaining global free lists and a heap-lock arbitration policy inspired by algorithms discussed at ACM SIGOPS meetings. Hoard’s global coordination is designed to satisfy provable bounds on memory blowup, a theoretical emphasis aligned with work from ETH Zurich and University of Cambridge on resource-usage guarantees.

Performance and Scalability

Hoard targets reduction of contention for the pthread allocator paths used in multithreaded applications common in environments like Apache HTTP Server and Nginx. Benchmarks compared Hoard against allocators developed in projects at Linux Foundation, FreeBSD, and NetBSD, and against allocator implementations evaluated by teams at Intel and AMD. Performance studies presented at EuroSys and SC Conference showed Hoard improving throughput on multicore systems produced by vendors such as Dell, HP, and Lenovo when running services developed by companies like Netflix and Twitter. Scalability tests used parallel workloads from suites by SPEC and research datasets from Google and Microsoft Azure, revealing lower lock contention and better cache-locality than legacy allocators documented by Sun Microsystems engineers. Comparative analyses in proceedings from USENIX ATC and ACM ICCD highlighted trade-offs between latency, fragmentation, and NUMA-aware placement, topics also explored by researchers at University of Illinois Urbana-Champaign and Rutgers University.

Implementation and Usage

Hoard’s reference implementation is written in C and C++, compatible with toolchains from GNU Compiler Collection and Clang/LLVM, and integrates with runtime environments used in projects from Red Hat and Canonical Ltd.. It provides hooks for runtime configuration used in server stacks such as Lighttpd and application platforms like Node.js and Tomcat. Deployments in research prototypes targeted multicore systems from ARM Holdings and IBM POWER architectures; development and debugging involved tools from Valgrind and GDB. The allocator’s BSD-like licensing facilitated inclusion in academic projects at Harvard University and open-source distributions maintained by communities around Debian and Arch Linux. Users incorporate Hoard into build systems using CMake and integration testing with continuous integration services popularized by Travis CI and Jenkins.

Comparisons and Alternatives

Hoard is frequently compared to allocators such as the Doug Lea malloc (dlmalloc), jemalloc, tcmalloc, and the default allocators maintained by glibc and FreeBSD developers. Evaluations by teams at Google Research and Mozilla Foundation contrast Hoard’s global coordination and fragmentation bounds with jemalloc’s arena-based design and tcmalloc’s thread-caching strategies developed at Google. Studies at Microsoft Research and Facebook weigh Hoard’s provable memory blowup guarantees against practical throughput optimizations in allocators from Facebook’s Folly library and allocator variants used in Windows NT runtime environments. Alternative approaches from academic groups at ETH Zurich, Princeton University, and UC Berkeley emphasize NUMA-aware allocation, lock-free freelists, and hardware transactional memory techniques.

Security and Reliability

Hoard’s structure affects reliability and mitigates certain classes of concurrency bugs that have been documented in advisories from CERT and incident analyses at US-CERT. Its segregation of per-thread heaps reduces cross-thread corruption risks noted in vulnerabilities assessed by teams at Microsoft Security Response Center and Google Project Zero, while fragmentation control relates to denial-of-service mitigation strategies discussed at Black Hat and DEF CON conferences. Integration with memory error detection tools from AddressSanitizer and runtime diagnostics used by Facebook and Mozilla aids in exposing use-after-free and double-free bugs investigated by researchers at Imperial College London and Georgia Institute of Technology. Hoard’s deterministic allocation policies have been used in reliability studies published in proceedings of IEEE Reliability Society events.

Category:Memory management