Check out my first novel, midnight's simulacra!
Libtorque
My project for Professor Rich Vuduc's Fall 2009 CSE6230, libtorque is a multithreaded event library for UNIX designed to take full advantage of the manycore, heterogenous, NUMA future. Previous, non-threaded event libraries include libevent, libev and liboop. My project proposal suggests motivation for libtorque: I believe it necessary to take scheduling and memory-placement decisions into account to most optimally handle events, especially on manycore machines and especially to handle unexpected traffic sets (denial of service attacks, oversubscribed pipes, mixed-latency connections, etc).
Resources
- git hosting from GitHub:
- Lots of good data in the README!
- Available from the dankamongmen/libtorque project page
- git clone from git://github.com/dankamongmen/libtorque.git
- bugzilla, hosted here on http://dank.qemfd.net/bugzilla/
Milestones
- 2009-11-19: CSE 6230 checkpoint
- 2009-12-10: CSE 6230 due date
Design/Functionality
libtorque exposes an affinity-managing, continuations-based, architecture-aware, multithreaded scheduler and various utility functions, including an architecture-, OS-, and thread-aware allocator with strong scheduler feedbacks. It can analyze arbitrary object code via libdl and libelf, discover where instructions live, and allocate around those areas. It can take dynamically balanced or (static) asymmetric interrupt loads into account. By making service decisions and allocations based on whole-system effects, libtorque provides low latency and high throughput under even pedantically asymmetric, irregular loads. By carefully distributing edge-triggered event descriptors among various processors' notification sets, highly scalable use is made of advanced low-level primitives such as kqueue and epoll. Through the use of lazy asynchronous I/O, each (expensive!) core is kept busy doing real work.
- Q: Why aren't you letting the OS manage scheduling, thus following the advice of just about everyone (assuming no real-time requirements)?
- A: Scalable event notification on UNIX is based around stateful polling (even in signal-driven POSIX asynchronous I/O, we want to use stateful polling on signals both for performance and event unification). The distribution of these events, at heart, drives scheduling. Since we must influence this distribution, we must make scheduling decisons.
System discovery
- Full support for CPUID as most recently defined by Intel and AMD (more advanced, as of 2009-10-31, than x86info)
- Full support for Linux and FreeBSD's native cpuset libraries, and SGI's libcpuset and libNUMA
- Discovers and makes available, for each processor type:
- ISA, ISA-specific capabilities, and number of concurrent threads supported (degrees of SMT)
- Line count, associativity, line length, geometry, and type of all caches
- Entry count, associativity, page size and type of all TLBs
- Inclusiveness relationships among cache and TLB levels
- Interconnection topology, APIC ID's, and how caches are shared among processors
- More: properties of hardware prefetching, ability to support non-temporal loads (MOVNTDQA, PREFETCHNTA, etc)
- Discovers and makes available, for each memory node type:
- Connected processor groups and relative distance information
- Number of pages and bank geometry
- More: OS page prefetching policy, error-recovery info
Scheduling
System architecture, allocations, continuation targets, event pattern and of course processing itself play a role in our scheduling. We first construct the scheduling universe, a scheduling model using the hardware detected during library initialization. Taken from src/libtorque/topology.h (commit a03fc592d03ccc9c20b016faaa1362b88bc07417):
// We are not considering distributed systems in this model. // // The scheduling universe is defined by an ordered set of $N > 1$ levels // $L_0..L_n$. A scheduling group is a structural isomorphism, a counter $C$ of // instances, and the $C$ affinity masks of corresponding processing elements. // A level contains $N > 0$ scheduling groups and a description of hardware // unique to that level. A unique surjection from classes of usable hardware to // levels (no hardware class is described at two levels, and all usable // hardware is described by the scheduling universe) is defined by via our // discovery algorithm. // // The reality compactifies things a bit, but in theory: // Level L_0 consists of scheduling groups over threads and processor types. // Each successive level $L_{n+1}, n >= 0$ extends $L_n$'s elements. For each // element $E \in L_n$, extend $E$ through those paths reachable at equal cost // from all elements in $E$. // // Examples: // A uniprocessor, no matter its memories, is a single topology level. // 2 unicore, unithread SMP processors with distinct L1 and shared L2 are two // topology levels: 2x{proc + L1} and {L2 + memory}. Split the L2, and they // remain two topology levels: 2x{proc + L1 + L2} and {memory}. Combine both // caches for a single level: {2proc + L1 + L2 + memory}. Sum over a level's // processors' threads for a thread count. // // Note that SMT does not come into play in shaping the topology hierarchy, // only in calculating the number of threads in a topological group. We only // schedule to SMT processors if // - shared data is allocated which fits entirely within the group's cache, or // - we need to.
Robustness
As the number of dies grow, so do interconnects. We're extending Moore's Law, but what about Weibull's distribution? As the number of cores rises, so too does the likelyhood that a processor or DIMM failure will be seen across the lifetime of a machine (remember, a failure distribution exponential as a function of time suggests only ~37% likelyhood of any given component reaching MTBF). Hot-pluggable processors and memories are also likely to be encountered, especially in cloud/virtual environments. For that matter, sysadmins can reconfigure cpusets and NUMA on the fly, or even migrate our process between sets. We ought design with all this in mind.
- Determine: how are Linux/FreeBSD processes notified/affected when their processors or memories fail?
- Determine: how to recover/redetect/redistribute?
References/Prior Art
- Philip Mucci's "Linux Multicore Performance Analysis and Optimization in a Nutshell", delivered at NOTUR 2009
- Elmeleegy et al's "Lazy Asynchronous I/O", USENIX 2004
- PGAS: Kathy Yelick's "Performance and Productivity Opportunities using Global Address Space Programming Models", 2006
- Emery Berger's Hoard and other manycore-capable allocators (libumem aka magazined slab, Google's ctmalloc, etc).
- Williams, Vuduc, Oliker, Shalf, Yelick, Demmel. "Tuning Sparse Matrix Vector Multiplication for multi-core SMPs", 2007
- A lot of discursive theorizing/ruminating is captured on my Fast UNIX Servers page
Phatty ASCII Logo
888 ,e, 888 d8 "...tear the roof off the sucka..." 888 " 888 88e d88 e88 88e 888,8, e88 888 8888 8888 ,e e, 888 888 888 888b d88888 d888 888b 888 " d888 888 8888 8888 d88 88b 888 888 888 888P 888 Y888 888P 888 Y888 888 Y888 888P 888 , 888 888 888 88" 888 "88 88" 888 "88 888 "88 88" "YeeP" _____________________________________________ 888 _________________ continuation-based unix i/o for manycore numa\888/© nick black 2009