Outline �� 0024 Spring 2010 � – 24 :: 2 �
Parallel application development � 0024 Spring 2010 � – 24 :: 3 �
0024 Spring 2010 � – 24 :: 4 �
Lock data, not code � 0024 Spring 2010 � – 24 :: 5 �
Do you really need locks? � No shared data => no need for locks � � � Recall that CSP gives you a model to avoid locks � � � No free lunch � Lock-free data structures � � � Mutex-free by design � � � Growing number of class/data structures � 0024 Spring 2010 � – 24 :: 6 �
Detour: no shared data � What if we could write programs so that there are no side-effects? � � � Think about the simple fi nite impulse response fi lter for N inputs � � � Think of computing an expensive function for N numbers � � � Think of searching for a string in N documents � 0024 Spring 2010 � – 24 :: 7 �
MapReduce � Basic idea: Parallel computing framework for restricted parallel programming model � Useful to distribute work to a farm (cluster) of compute nodes � User speci fi es what needs to be done for each data item (“map”) and how results are to be combined (“reduce”) � Libraries take care of everything else � � � Parallelization � � � Fault Tolerance � � � Data Distribution � � � Load Balancing � 0024 Spring 2010 � – 24 :: 8 �
MapReduce � Map() �� � � Process a key/value pair to generate intermediate key/value pairs � Reduce() �� � � Merge all intermediate values associated with the same key � Names originated in the functional programming world … but slightly different semantics � 0024 Spring 2010 � – 24 :: 9 �
Example: Counting Words � Map() �� � � Input < fi lename, fi le text> � � � Parses fi le and emits <word, count> pairs � � � E.g. <”hello”, 1> � Reduce() �� � � Sums all values for the same key � � � Emits <word, TotalCount> � � � E.g. <”hello”, 5 > <”hello”, 1> � <”hello”, 2 > <”hello”, 7 > => <”hello”, 15> � 0024 Spring 2010 � – 24 :: 10 �
Example Use of MapReduce � Counting words in a large set of documents � map(string key, string value) � //key: document name //value: document contents for each word w in value EmitIntermediate(w, “1”); reduce(string key, iterator values) � //key: word //values: list of counts int results = 0; for each v in values result += ParseInt(v); Emit(AsString(result)); � 0024 Spring 2010 � – 24 :: 11 �
Data Distribution � Input fi les are split into pieces � � � distributed fi le system � Intermediate fi les created from map tasks are written to local disk � Output fi les are written to distributed fi le system � 0024 Spring 2010 � – 24 :: 12 �
Assigning Tasks � Many copies of user program are started � Tries to utilize data localization by running map tasks on machines with data � One instance becomes the master � Master fi nds idle machines and assigns tasks � 0024 Spring 2010 � – 24 :: 13 �
0024 Spring 2010 � – 24 :: 14 �
MapReduce � 0024 Spring 2010 � – 24 :: 15 �
Do you really need locks? � No shared data => no need for locks � � � Recall that CSP gives you a model to avoid locks � � � No free lunch � Lock-free data structures � � � Mutex-free by design � � � Growing number of class/data structures � 0024 Spring 2010 � – 24 :: 16 �
Why Locking Doesn � t Scale � Not Robust � Relies on conventions � Hard to Use � � � Conservative � � � Deadlocks � � � Lost wake-ups � Not Composable � 0024 Spring 2010 � – 24 :: 17 �
Locks are not Robust � If a thread holding a lock is delayed … No one else can make progress 0024 Spring 2010 � – 24 :: 18 �
Why Locking Doesn � t Scale � Not Robust � Relies on conventions � Hard to Use � � � Conservative � � � Deadlocks � � � Lost wake-ups � Not Composable � 0024 Spring 2010 � – 24 :: 19 �
Locking Relies on Conventions � Relation between � � � Lock bit and object bits � � � Exists only in programmer � s mind � Actual comment from Linux Kernel (hat tip: Bradley Kuszmaul) /* * When a locked buffer is visible to the I/O layer * BH_Launder is set. This means before unlocking * we must clear BH_Launder,mb() on alpha and then * clear BH_Lock, so no reader can see BH_Launder set * on an unlocked buffer and then risk to deadlock. */ 0024 Spring 2010 � – 24 :: 20 �
Why Locking Doesn � t Scale � Not Robust � Relies on conventions � Hard to Use � � � Conservative � � � Deadlocks � � � Lost wake-ups � Not Composable � 0024 Spring 2010 � – 24 :: 21 �
Sadistic Homework � Fifo queue enq( y) enq( x) No interference if ends “far enough” apart 0024 Spring 2010 � – 24 :: 22 �
Sadistic Homework � Double-ended queue deq( ) deq( ) Interference OK if ends “close enough” together 0024 Spring 2010 � – 24 :: 23 �
Sadistic Homework � Double-ended queue deq( ) deq( ) � Make sure suspended dequeuers awake as needed 0024 Spring 2010 � – 24 :: 24 �
In Search of the Lost Wake-Up � Waiting thread doesn � t realize when to wake up � It � s a real problem in big systems � � � “Calling pthread_cond_signal() or pthread_cond_broadcast() when the thread does not hold the mutex lock associated with the condition can lead to lost wake-up bugs.” � from Google™ search for “lost wake-up” 0024 Spring 2010 � – 24 :: 25 �
You Try It … � One lock? � � � Too Conservative � Locks at each end? � � � Deadlock, too complicated, etc � Waking blocked dequeuers? � � � Harder than it looks � 0024 Spring 2010 � – 24 :: 26 �
Actual Solution � Clean solution would be a publishable result � [Michael & Scott, PODC 96] � High performance fi ne-grained lock-based solutions are good for libraries… � not general consumption by programmers � 0024 Spring 2010 � – 24 :: 27 �
Why Locking Doesn � t Scale � Not Robust � Relies on conventions � Hard to Use � � � Conservative � � � Deadlocks � � � Lost wake-ups � Not Composable � 0024 Spring 2010 � – 24 :: 28 �
Locks do not compose � Hash Table lock T1 lock T 1 Must lock T 1 before adding add(T 1 , item) item item Move from T 1 to T 2 lock T1 lock T 1 lock T2 lock T 2 Must lock T 2 delete(T 1 , item) before deleting add(T 2 , item) item item from T 1 Exposing lock internals breaks abstraction 0024 Spring 2010 � – 24 :: 29 �
Monitor Wait and Signal � Empty buffer zzz Yes! If buffer is empty, wait for item to show up 0024 Spring 2010 � – 24 :: 30 �
Wait and Signal do not Compose � empty Wait for either? empty zzz… 0024 Spring 2010 � – 24 :: 31 �
The Transactional Manifesto � What we do now is inadequate to meet the multi-core challenge � Research Agenda � � � Replace locking with a transactional API � � � Design languages to support this model � � � Implement the run-time to be fast enough � 0024 Spring 2010 � – 24 :: 32 �
Transactions � Atomic � � � Commit: takes effect � � � Abort: effects rolled back � � � Usually retried � Linearizable � � � Appear to happen in one-at-a-time order � 0024 Spring 2010 � – 24 :: 33 �
Atomic Blocks � at om i c { x. r em ove( 3) ; y. add( 3) ; } at om i c { y = nul l ; } 0024 Spring 2010 � – 24 :: 34 �
Atomic Blocks � at om i c { No data race x. r em ove( 3) ; y. add( 3) ; } at om i c { y = nul l ; } 0024 Spring 2010 � – 24 :: 35 �
Sadistic Homework Revisited � Publ i c voi d Lef t Enq( i t em x) { Q node q = new Q node( x) ; q. l ef t = t hi s. l ef t ; t hi s. l ef t . r i ght = q; t hi s. l ef t = q; } Write sequential code (1) 0024 Spring 2010 � – 24 :: 36 �
Sadistic Homework Revisited � Publ i c voi d Lef t Enq( i t em x) { at om i c { Q node q = new Q node( x) ; q. l ef t = t hi s. l ef t ; t hi s. l ef t . r i ght = q; t hi s. l ef t = q; } } (1) 0024 Spring 2010 � – 24 :: 37 �
Sadistic Homework Revisited � Publ i c voi d Lef t Enq( i t em x) { at om i c { Q node q = new Q node( x) ; q. l ef t = t hi s. l ef t ; t hi s. l ef t . r i ght = q; t hi s. l ef t = q; } } Enclose in atomic block (1) 0024 Spring 2010 � – 24 :: 38 �
Warning � Not always this simple � � � Conditional waits � � � Enhanced concurrency � � � Overlapping locks � But often it is � � � Works for sadistic homework � 0024 Spring 2010 � – 24 :: 39 �
Composition � Publ i c voi d Tr ansf er ( Q ueue q1, q2) { at om i c { O bj ect x = q1. deq( ) ; q2. enq( x) ; } } Trivial or what? 0024 Spring 2010 � – 24 :: 40 �
Wake-ups: lost and found � Publ i c O bj ect Lef t Deq( ) { at om i c { i f ( t hi s. l ef t == nul l ) r et r y; … } } Roll back transaction and restart when something changes 0024 Spring 2010 � – 24 :: 41 �
Recommend
More recommend
