Dark Silicon, Mobile Devices, and Possible Open-Source Solutions
•
5 likes•2,739 views
This document summarizes a presentation about dark silicon in mobile devices and possible open source solutions. It discusses how power and thermal constraints are more severe for mobile devices due to limited battery progress and no fans. It also covers big.LITTLE scheduling, thread-level parallelism challenges, and user-level threading libraries like AsyncTask. Finally, it notes that while some open source parallel programming frameworks exist, fully utilizing parallelism on mobile and addressing dark silicon remain challenges with no widely adopted solutions.
Recommended
More Related Content
What's hot
What's hot (20)
Similar to Dark Silicon, Mobile Devices, and Possible Open-Source Solutions
Similar to Dark Silicon, Mobile Devices, and Possible Open-Source Solutions (20)
More from Koan-Sin Tan
More from Koan-Sin Tan (8)
Recently uploaded
Recently uploaded (20)
Dark Silicon, Mobile Devices, and Possible Open-Source Solutions
- 1. Dark Silicon, Mobile Devices, and Possible Open Source Solutions Koan-Sin Tan freedom@computer.org COSCUP 2013, Aug. 3rd,TICC,Taipei Friday, August 23, 13
- 2. • Software engineer, veteran open-source user • Learned something about light-weight process (LWP) on Sun OS 4.x in early 1990s • Did a user-level thread library on 386BSD with a classmate in 1992 • Was involved in big.LITTLE scheduling work recently Friday, August 23, 13
- 3. Samsung “optimization” for senchmarks http://www.anandtech.com/show/7187/looking-at- cpugpu-benchmark-optimizations-galaxy-s-4 Friday, August 23, 13
- 4. Friday, August 23, 13
- 6. • “Dark Silicon refers to the exponentially increasing number of a chip's transistors that must remain passive, or "dark", in order to stay within a chip's power budget” Friday, August 23, 13
- 7. Figure from the textbook. We know we are in CMP era. “Since 2003, the limits of power and available instruction- level parallelism have slowed uniprocessor performance.” Friday, August 23, 13
- 8. Dennard scaling hits the wall • Dennard Scaling • When voltages are scaled along with all dimensions, a device’s electric fields remain constant, and most device characteristics are preserved • scaling maintains constant power density • logic area and power is scaled down by alpha^2 • energy per transition is scaled down by alpha^3, but frequency is scaled up by 1/alpha, resulting in an alpha^2 decrease in power per gate • ........ • google Dennard Scaling you can find more information, such as, http:// www1.cs.columbia.edu/~cs4824/lectures/csee4824_f12_lec22.pdf Friday, August 23, 13
- 9. Mobile Devices • Both power and thermal constrains are more severe than desktop devices • The progress of battery is relatively slow • You don’t want put a fan into you smartphone • conduction, convection, radiation Friday, August 23, 13
- 10. Yes, modern high-end mobile processors have serious thermal problems.Tegra 4 game console figure from iFixit Friday, August 23, 13
- 11. Nexus 10 Thermal Throttling • Antutu 3.0.2 • Unit for X axis is 200 ms • It reaches 80 ˚C in 20 second • Throttling starts at 80 ˚C; stops at 78 ˚C • Throttling is to decrement themaximum freq value of cpufreq Friday, August 23, 13
- 14. Introducingbig.LITTLE Figure 28-3 Processor DVFS curves In a big.LITTLE system these operating points are applied both to the Cortex-A15 and Cortex-A7 processors. When the Cortex-A7 processor is executing the OS can tune the operating points as it would for an existing platform with a single applications processor. When the Cortex-A7 processor is at its highest operating point (Figure 28-3), if more performance is required a switch is invoked that transfers the OS and applications to the Cortex-A15 processor. Further DVFS tuning takes place on the Cortex-A15 processor if required, as the operating load increases. Migration requires rapid context switching capability. Coherency is clearly a critical enabler in achieving a fast task migration time as it allows the state that has been saved on the outbound (migrated from) processor to be snooped and restored on the inbound (migrated to) processor rather than going via main memory. Additionally, for Cluster migration, (or for CPU migration when all processors have been switched) because the L2 cache of the outbound processor is coherent it can remain powered up after a task migration to improve the cache warming time of ARM big.LITTLE Friday, August 23, 13
- 15. Thread-Level Parallelism • Thread-level Parallelism (TLP) is an index you can treat it as number of threads running concurrently • a table from an ISCA ‘10 paper named “Evolution of thread-level parallelism in desktop applications” • 2000, 2010 • mobile devices are worse • http://dl.acm.org/citation.cfm? id=1816000 Friday, August 23, 13
- 16. Parallel Programming Could Help a Bit • Parallel computing/programming has been there for a long time • You know pthread and OpenMP are available and C++11 came with currency support • Java use thread and its synchronization model • “Why Threads Are A Bad Idea”, by John Ousterhout, http://www.cc.gatech.edu/ classes/AY2009/cs4210_fall/papers/ousterhout-threads.pdf • Thread is “easy: to describe; to use; to get wrong” to quote Andrew Birrell, http://www.cs.princeton.edu/courses/archive/spr07/cos598A/lectures/ Birrell.pdf • For more theoretical explanation, see “The Problems with Threads” by Edward Lee, http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-1.pdf • And you know that except shared memory model, there is message passing computing model. And more, e.g., actors, data-flow, systolic array, etc. Friday, August 23, 13
- 17. Threads are Bad Ideas? • “Why Threads Are Bad Ideas”, John Ousterhout, 1995, http:// www.cc.gatech.edu/classes/AY2009/ cs4210_fall/papers/ousterhout- threads.pdf • Yes, It’s a bit dated. Some of those points are no longer valid; many of them stand the test of time • Threads: • Too hard for most programmers to use • Even for experts, development is painful Friday, August 23, 13
- 18. Some of Ousterhout’s arguments remain valid • Synchronization • manually set of mutex/lock • deadlock: yes deadlock • hard to debug • threads breaks modularization • callbacks don’t work with locks Friday, August 23, 13
- 19. thread is easy to get wrong • Manual selection of mutual exclusion: • Default is too little (and hence races) • Easy fix is too much (deadlocks or blank stares) • Projects don’t create hierarchical abstractions • Can’t decide and/or maintain acyclic locking order • “Composition” requires entire new abstractions • “Clever” optimizations aren’t maintainable • ..... Friday, August 23, 13
- 20. User-level libraries, frameworks • Android AsyncTask • a class to help perform background operations and publish results on the UI thread without having to manipulate threads and/or handlers • http://developer.android.com/reference/android/os/AsyncTask.html • Intel Threading Building Blocks (TBB) • http://threadingbuildingblocks.org/, http://en.wikipedia.org/wiki/ Intel_Threading_Building_Blocks • works on Android x86 and ARM • Apple Grand Central Dispatch (GCD) • http://developer.apple.com/library/ios/#documentation/Performance/ Reference/GCD_libdispatch_Ref/ • Software Transactional Memory • http://gcc.gnu.org/wiki/TransactionalMemory Friday, August 23, 13
- 21. Language extension • Intel Cilk Plus • http://cilkplus.org/, http://en.wikipedia.org/ wiki/Intel_Cilk_Plus • open sourced, trying to get into gcc and llvm • Apple blocks • http://developer.apple.com/library/ios/ #documentation/cocoa/Conceptual/Blocks/ Friday, August 23, 13
- 22. OpenCL Related • OpenCL • pocl, http://pocl.sourceforge.net/ • OpenCL and Java • Aparapi, https://code.google.com/p/aparapi/ • Smuatra, http://openjdk.java.net/projects/sumatra/ • RenderScript • in AOSP • ThorScript • will be open-sourced Friday, August 23, 13
- 23. Cilk Plus: simple language extensions originated from Charles Leiserson Friday, August 23, 13
- 24. Simple Cilk Plus Example int fib(int n) { if (n < 2) return n; int x = fib(n-1); int y = fib(n-2); return x + y; } int fib(int n) { if (n < 2) return n; int x = clik_spawn fib(n-1); int y = fib(n-2); cilk_sync; return x + y; } Friday, August 23, 13
- 25. simple GCD+blocks dispatch_group_t group = dispatch_group_create(); fib = ^() { if (n < 2) { result = n; return; } __block int x, y; int m = n; n = m - 1; dispatch_group_async(group, a_queue, ^{fib(); x = result;}); dispatch_group_wait(group, DISPATCH_TIME_FOREVER); n = m - 2; dispatch_sync(a_queue, ^{fib(); y = result;}); n = m; result = x + y; return; }; Friday, August 23, 13
- 26. data parallel fib() looks more reasonable int fib(int n) { if (n < 2) return n; int p = 0, q = 1, result =0; cilk_for (int i=2; i <= n; i++) { result = p + q; p = q; q = result; } return result; } TextText Text n.b.: in case you didn’t notice, this may produce wrong results because of loop-carried dependency Friday, August 23, 13
- 27. parallel fib() with GCD and blocks int(^fib)(int); fib = ^(int n){ if (n < 2) return n; __block int p = 0, q = 1, result = 0; dispatch_apply(n-1, dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0), ^(size_t i) { result = p + q; p = q; q = result; }); return result; }; Friday, August 23, 13
- 28. GCD is can be used with OpenCL And GCD • That’s what is available on Mac OS X and iOS • Nope, iOS didn’t open OpenCL yet. But you can find how to use OpenCL for ARM on iOS easily Friday, August 23, 13
- 29. What are available • Task-parallel and data-parallel constructs, libraries or languguages • Lambda, closure, continuation, etc. • Queue, queue management: load balance, work stealing, etc • Data structures, e.g.,TBB • Lock-less synchronization Friday, August 23, 13
- 30. Lockfree synchronization • In case you didn’t know it, NO, it’s not new at all • Linux has been used RCU (Read-Copy- Update) for several years • In fact, it’s there since 1970s, see Kung’s 1980 paper proposed RCU-like mechanism. Friday, August 23, 13
- 31. Kernel • big.LITTLE • IKS: in-kernel-switcher • related code being upstreaming after 3.10 • Global Task Scheduling (GTS), Heterogenous Multi-Processor (HMP) • Current CFS maintainer Ingo didn’t like GTS’s power-saving part • Power Management • So many mechanisms: cpufreq, cpuidle, runtime PM, CCF, etc. • Linaro has a wiki page on how to/what to enable/implement for a new SoC • Thermal Management • Throttling, e.g., ask related components to slow down so that less heat will be generated Friday, August 23, 13
- 32. Linaro In-kernel Switcher Friday, August 23, 13
- 33. Global Task-Scheduling (GTS) Friday, August 23, 13
- 34. Many are remained to be done • No widely used open-source power or thermal management framework available? • Some problems are fundamental hard to parallelized, e.g., • parsing in browser: nowadays, webkit and firefox use LALR(1) or similar parsing algorithm • No full-featured open-source OpenCL implementation for GPGPU Friday, August 23, 13
- 35. Wrap-up • “dark silicon” is reality on mobile devices, • power wall and thermal wall • parallel/concurrent code isn’t popular on mobile devices (yet) • discussed some possible free and open source solutions • many remained to be done Friday, August 23, 13