Is Parallel Programming Hard, And, If So, What Can You Do About It?

More documents

Recommendations

Info

$TeX op Mac OS X, met teTeX en TeXShop - Nluug$

196 APPENDIX D. READ-COPY UPDATE IMPLEMENTATIONSonlyiftheaforementioneddyntickscounteriseven.In other words, NMI’s refrain from manipulating thedynticks_nmi counter if the NMI occurred in nondynticks-idlemode or within an interrupt handler.The only difference between these two functions istheerrorchecks,asrcu_nmi_enter()mustleavethedynticks_nmi counter with an odd value, and rcu_nmi_exit() must leave this counter with an evenvalue.D.2.7.7 Note That a CPU is in Dynticks IdleModeThe force_quiescent_state() function implementsa three-phase state machine. The first phase(RCU_INITIALIZING) waits for rcu_start_gp() tocomplete grace-period initialization. This stateis not exited by force_quiescent_state(), butrather by rcu_start_gp().In the second phase (RCU_SAVE_DYNTICK), thedyntick_save_progress_counter() functionscans the CPUs that have not yet reported aquiescent state, recording their per-CPU dynticksand dynticks_nmi counters. If these counters bothhave even-numbered values, then the correspondingCPU is in dynticks-idle state, which is thereforenoted as an extended quiescent state (reported viacpu_quiet_msk()).In the third phase (RCU_FORCE_QS), the rcu_implicit_dynticks_qs() function again scans theCPUs that have not yet reported a quiescent state(either explicitly or implicitly during the RCU_SAVE_DYNTICK phase), again checking the per-CPUdynticks and dynticks_nmi counters. If each ofthesehaseitherchangedinvalueorisnoweven, thenthe corresponding CPU has either passed through oris now in dynticks idle, which as before is noted asan extended quiescent state.If rcu_implicit_dynticks_qs() finds that agiven CPU has neither been in dynticks idle modenor reported a quiescent state, it invokes rcu_implicit_offline_qs(), whichcheckstoseeifthatCPU is offline, which is also reported as an extendedquiescent state. If the CPU is online, thenrcu_implicit_offline_qs() sends it a rescheduleIPI in an attempt to remind it of its duty to reporta quiescent state to RCU.Note that force_quiescent_state() does notdirectly invoke either dyntick_save_progress_counter() or rcu_implicit_dynticks_qs(), insteadpassing these functions to an intervening rcu_process_dyntick() function that abstracts out thecommon code involved in scanning the CPUs andreporting extended quiescent states.Quick Quiz D.13: And what happens if oneCPU comes out of dyntick-idle mode and thenpassed through a quiescent state just as anotherCPU notices that the first CPU was in dyntick-idlemode? Couldn’t they both attempt to report a quiescentstateatthesametime,resultinginconfusion?Quick Quiz D.14: But what if all the CPUs endup in dyntick-idle mode? Wouldn’t that prevent thecurrent RCU grace period from ever ending?Quick Quiz D.15: Given that force_quiescent_state() is a three-phase state machine,don’t we have triple the scheduling latency due toscanning all the CPUs?D.2.7.8 Offline a CPUCPU-offline events cause rcu_cpu_notify() to invokercu_offline_cpu(), which in turn invokes__rcu_offline_cpu() on both the rcu and thercu bh instances of the data structures. This functionclears the outgoing CPU’s bits so that futuregrace periods will not expect this CPU to announcequiescentstates, andfurtherinvokescpu_quiet()inorder to announce the offline-induced extended quiescentstate. This work is performed with the global->onofflock held in order to prevent interferencewith concurrent grace-period initialization.Quick Quiz D.16: But the other reason to hold->onofflock is to prevent multiple concurrent online/offlineoperations, right?D.2.7.9 Online a CPUCPU-online events cause rcu_cpu_notify() to invokercu_online_cpu(), which initializes the incomingCPU’s dynticks state, and then invokesrcu_init_percpu_data() to initialize the incomingCPU’s rcu_data structure, and also to setthis CPU’s bits (again protected by the global ->onofflock) so that future grace periods will waitfor a quiescent state from this CPU. Finally, rcu_online_cpu() sets up the RCU softirq vector forthis CPU.Quick Quiz D.17: Givenalltheseacquisitionsoftheglobal->onofflock, won’ttherebehorriblelockcontention when running with thousands of CPUs?Quick Quiz D.18: Why not simplify the codeby merging the detection of dyntick-idle CPUs withthat of offline CPUs?
D.2. HIERARCHICAL RCU OVERVIEW 197D.2.7.10 Detect a Too-Long Grace PeriodWhen the CONFIG_RCU_CPU_STALL_DETECTOR kernelparameter is specified, the record_gp_stall_check_time() function records the time and also atimestamp set three seconds into the future. If thecurrentgraceperiodstillhasnotendedbythattime,the check_cpu_stall() function will check for theculprit, invoking print_cpu_stall() if the currentCPU is the holdout, or print_other_cpu_stall()if it is some other CPU. A two-jiffies offset helps ensurethat CPUs report on themselves when possible,taking advantage of the fact that a CPU can normallydo a better job of tracing its own stack thanit can tracing some other CPU’s stack.D.2.8 TestingRCU is fundamental synchronization code, so anyfailure of RCU results in random, difficult-to-debugmemory corruption. It is therefore extremely importantthat RCU be highly reliable. Some of thisreliability stems from careful design, but at the endof the day we must also rely on heavy stress testing,otherwise known as torture.Fortunately, although there has been some debateas to exactly what populations are covered by theprovisions of the Geneva Convention it is still thecase that it does not apply to software. Therefore,it is still legal to torture your software. In fact, itis strongly encouraged, because if you don’t tortureyour software, it will end up torturing you by crashingat the most inconvenient times imaginable.Therefore, we torture RCU quite vigorously usingthe rcutorture module.However, it is not sufficient to torture thecommon-case uses of RCU. It is also necessary totorture it in unusual situations, for example, whenconcurrently onlining and offlining CPUs and whenCPUs are concurrently entering and exiting dynticksidle mode. I use a script @@@ move to CodeSamples,ref@@@andusethetest_no_idle_hzmoduleparameter to rcutorture to stress-test dynticks idlemode. Just to be fully paranoid, I sometimes run akernbench workload in parallel as well. Ten hoursof this sort of torture on a 128-way machine seemssufficient to shake out most bugs.Eventhisisnotthecompletestory. AsAlexeyDobriyanand Nick Piggin demonstrated in early 2008,it is also necessary to torture RCU with all relevantcombinations of kernel parameters. The relevantkernel parameters may be identified using yetanother script@@@ move toCodeSamples, ref@@@1. CONFIG_CLASSIC_RCU: Classic RCU.2. CONFIG_PREEMPT_RCU: Preemptable (real-time)RCU.3. CONFIG_TREE_RCU: Classic RCU for huge SMPsystems.4. CONFIG_RCU_FANOUT: Number of children foreach rcu_node.5. CONFIG_RCU_FANOUT_EXACT: Balance the rcu_node tree.6. CONFIG_HOTPLUG_CPU: Allow CPUs to be offlinedand onlined.7. CONFIG_NO_HZ: Enable dyntick-idle mode.8. CONFIG_SMP: Enable multi-CPU operation.9. CONFIG_RCU_CPU_STALL_DETECTOR: EnableRCU to detect when CPUs go on extendedquiescent-state vacations.10. CONFIG_RCU_TRACE: Generate RCU trace filesin debugfs.We ignore the CONFIG_DEBUG_LOCK_ALLOC configurationvariable under the perhaps-naive assumptionthat hierarchical RCU could not have brokenlockdep. There are still 10 configuration variables,which would result in 1,024 combinations if theywere independent boolean variables. Fortunatelythe first three are mutually exclusive, which reducesthe number of combinations down to 384, butCONFIG_RCU_FANOUT can take on values from 2 to64, increasingthe number of combinations to12,096.This is an infeasible number of combinations.One key observation is that only CONFIG_NO_HZ and CONFIG_PREEMPT can be expected to havechanged behavior if either CONFIG_CLASSIC_RCU orCONFIG_PREEMPT_RCUareineffect,asonlytheseportionsof the two pre-existing RCU implementationswere changed during this effort. This cuts out almosttwo thirds of the possible combinations.Furthermore, not all of the possible values ofCONFIG_RCU_FANOUT produce significantly differentresults, in fact only a few cases really need to betested separately:1. Single-node “tree”.2. Two-level balanced tree.3. Three-level balanced tree.4. Autobalanced tree, where CONFIG_RCU_FANOUTspecifies an unbalanced tree, but such that itis auto-balanced in absence of CONFIG_RCU_FANOUT_EXACT.
Page 1 and 2:
Is Parallel Programming Hard, And,
Page 3 and 4:
Contents1 Introduction 11.1 Histori
Page 5 and 6:
CONTENTSv6 Locking 676.1 Staying Al
Page 7 and 8:
CONTENTSviiB Synchronization Primit
Page 9 and 10:
CONTENTSixE.7.1 Introduction to Pre
Page 11 and 12:
PrefaceThe purpose of this book is
Page 13 and 14:
Chapter 1IntroductionParallel progr
Page 15 and 16:
1.2. PARALLEL PROGRAMMING GOALS 3CP
Page 17 and 18:
1.3. ALTERNATIVES TO PARALLEL PROGR
Page 19 and 20:
1.4. WHAT MAKES PARALLEL PROGRAMMIN
Page 21 and 22:
1.5. GUIDE TO THIS BOOK 9other hand
Page 23 and 24:
Chapter 2Hardware and its HabitsMos
Page 25:
2.1. OVERVIEW 13Therefore, as shown
Page 28 and 29:
16 CHAPTER 2. HARDWARE AND ITS HABI
Page 30 and 31:
18 CHAPTER 2. HARDWARE AND ITS HABI
Page 32 and 33:
20 CHAPTER 3. TOOLS OF THE TRADE1 p
Page 34 and 35:
22 CHAPTER 3. TOOLS OF THE TRADE1 p
Page 36 and 37:
24 CHAPTER 3. TOOLS OF THE TRADE1.1
Page 38 and 39:
26 CHAPTER 3. TOOLS OF THE TRADEQui
Page 40 and 41:
28 CHAPTER 3. TOOLS OF THE TRADE
Page 42 and 43:
30 CHAPTER 4. COUNTING1 atomic_t co
Page 44 and 45:
32 CHAPTER 4. COUNTING4.2.3 Eventua
Page 46 and 47:
34 CHAPTER 4. COUNTINGvanish when t
Page 48 and 49:
36 CHAPTER 4. COUNTINGper-thread va
Page 50 and 51:
38 CHAPTER 4. COUNTING1 unsigned lo
Page 52 and 53:
Page 54 and 55:
42 CHAPTER 4. COUNTING1 #define THE
Page 56 and 57:
Page 58 and 59:
46 CHAPTER 4. COUNTINGReadsAlgorith
Page 60 and 61:
48 CHAPTER 5. PARTITIONING AND SYNC
Page 62 and 63:
Page 64 and 65:
Page 66 and 67:
Page 68 and 69:
Page 70 and 71:
Page 72 and 73:
Page 74 and 75:
Page 76 and 77:
Page 78 and 79:
Page 80 and 81:
68 CHAPTER 6. LOCKING1 int delete(i
Page 82 and 83:
70 CHAPTER 7. DATA OWNERSHIP
Page 84 and 85:
72 CHAPTER 8. DEFERRED PROCESSINGfo
Page 86 and 87:
74 CHAPTER 8. DEFERRED PROCESSINGth
Page 88 and 89:
76 CHAPTER 8. DEFERRED PROCESSING
Page 90 and 91:
78 CHAPTER 8. DEFERRED PROCESSINGfi
Page 92 and 93:
80 CHAPTER 8. DEFERRED PROCESSINGti
Page 94 and 95:
82 CHAPTER 8. DEFERRED PROCESSINGNo
Page 96 and 97:
84 CHAPTER 8. DEFERRED PROCESSING12
Page 98 and 99:
Page 100 and 101:
88 CHAPTER 8. DEFERRED PROCESSINGvo
Page 102 and 103:
90 CHAPTER 8. DEFERRED PROCESSINGLi
Page 104 and 105:
92 CHAPTER 8. DEFERRED PROCESSINGpe
Page 106 and 107:
94 CHAPTER 8. DEFERRED PROCESSINGCa
Page 108 and 109:
96 CHAPTER 8. DEFERRED PROCESSINGTh
Page 110 and 111:
Page 112 and 113:
Page 114 and 115:
Page 116 and 117:
104 CHAPTER 8. DEFERRED PROCESSINGs
Page 118 and 119:
106 CHAPTER 8. DEFERRED PROCESSINGo
Page 120 and 121:
108 CHAPTER 8. DEFERRED PROCESSING
Page 122 and 123:
110 CHAPTER 9. APPLYING RCU1 struct
Page 124 and 125:
112 CHAPTER 9. APPLYING RCU
Page 126 and 127:
114 CHAPTER 10. VALIDATION: DEBUGGI
Page 128 and 129:
116 CHAPTER 11. DATA STRUCTURES
Page 130 and 131:
118 CHAPTER 12. ADVANCED SYNCHRONIZ
Page 132 and 133:
Page 134 and 135:
Page 136 and 137:
Page 138 and 139:
Page 140 and 141:
Page 142 and 143:
Page 144 and 145:
Page 146 and 147:
Page 148 and 149:
Page 150 and 151:
138 CHAPTER 13. EASE OF USEFigure 1
Page 152 and 153:
140 CHAPTER 13. EASE OF USE
Page 154 and 155:
142 CHAPTER 14. TIME MANAGEMENT
Page 156 and 157:
144 CHAPTER 15. CONFLICTING VISIONS
Page 158 and 159: 146 CHAPTER 15. CONFLICTING VISIONS
Page 166 and 167: 154 APPENDIX A. IMPORTANT QUESTIONS
Page 168 and 169: 156 APPENDIX A. IMPORTANT QUESTIONS
Page 170 and 171: 158 APPENDIX B. SYNCHRONIZATION PRI
Page 172 and 173: 160 APPENDIX B. SYNCHRONIZATION PRI
Page 174 and 175: 162 APPENDIX C. WHY MEMORY BARRIERS
Page 196 and 197: 184 APPENDIX D. READ-COPY UPDATE IM
Page 256 and 257: 244 APPENDIX E. FORMAL VERIFICATION
Page 258 and 259:
246 APPENDIX E. FORMAL VERIFICATION
Page 260 and 261:
Page 262 and 263:
Page 264 and 265:
Page 266 and 267:
Page 268 and 269:
Page 270 and 271:
Page 272 and 273:
Page 274 and 275:
Page 276 and 277:
Page 278 and 279:
Page 280 and 281:
Page 282 and 283:
Page 284 and 285:
272 APPENDIX F. ANSWERS TO QUICK QU
Page 286 and 287:
Page 288 and 289:
Page 290 and 291:
Page 292 and 293:
Page 294 and 295:
Page 296 and 297:
Page 298 and 299:
Page 300 and 301:
Page 302 and 303:
Page 304 and 305:
Page 306 and 307:
Page 308 and 309:
Page 310 and 311:
Page 312 and 313:
Page 314 and 315:
Page 316 and 317:
Page 318 and 319:
Page 320 and 321:
Page 322 and 323:
Page 324 and 325:
Page 326 and 327:
Page 328 and 329:
Page 330 and 331:
Page 332 and 333:
Page 334 and 335:
Page 336 and 337:
Page 338 and 339:
Page 340 and 341:
Page 342 and 343:
330 APPENDIX G. GLOSSARY(2) A physi
Page 344 and 345:
332 APPENDIX G. GLOSSARYnear by. Th
Page 346 and 347:
334 APPENDIX G. GLOSSARY
Page 348 and 349:
336 BIBLIOGRAPHY[But97]USA, March 2
Page 350 and 351:
338 BIBLIOGRAPHY[HMB06][Hol03][HP95
Page 352 and 353:
340 BIBLIOGRAPHY[McK06] Paul E. McK
Page 354 and 355:
342 BIBLIOGRAPHYtor. Software - Pra
Page 356 and 357:
344 BIBLIOGRAPHY[UoC08][VGS08]Berke
Page 358:
346 APPENDIX H. CREDITSH.4 Original
show all

Is Parallel Programming Hard, And, If So, What Can You Do About It?

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?