DOCSLIB.ORG

On the Coexistence of Shared-Memory and Message-Passing in The

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
On the Coexistence of Shared-Memory and Message-Passing in The On the Co existence of SharedMemory and MessagePassing in the Programming of Parallel Applications 1 2 J Cordsen and W SchroderPreikschat GMD FIRST Rudower Chaussee D Berlin Germany University of Potsdam Am Neuen Palais D Potsdam Germany Abstract Interop erability in nonsequential applications requires com munication to exchange information using either the sharedmemory or messagepassing paradigm In the past the communication paradigm in use was determined through the architecture of the underlying comput ing platform Sharedmemory computing systems were programmed to use sharedmemory communication whereas distributedmemory archi tectures were running applications communicating via messagepassing Current trends in the architecture of parallel machines are based on sharedmemory and distributedmemory For scalable parallel applica tions in order to maintain transparency and eciency b oth communi cation paradigms have to co exist Users should not b e obliged to know when to use which of the two paradigms On the other hand the user should b e able to exploit either of the paradigms directly in order to achieve the b est p ossible solution The pap er presents the VOTE communication supp ort system VOTE provides co existent implementations of sharedmemory and message passing communication Applications can change the communication paradigm dynamically at runtime thus are able to employ the under lying computing system in the most convenient and applicationoriented way The presented case study and detailed p erformance analysis un derpins the applicability of the VOTE concepts for highp erformance parallel computing Intro duction Future parallel machines will exhibit a hybrid architecture based on shared memory and distributedmemory This line is already followed with all stateof theart sp ecial purp ose parallel computers Similar holds for parallel machines based on workstation clusters actual highp erformance workstations are shared memory multipro cessor systems Beside the diculty developing scalable and ecient parallel programs at all the problem of an application programmer is exploiting the various com puter architectures in the most ecient way Due to false sharing relying on the sharedmemory paradigm as intro duced by a virtual sharedmemory VSM system results in a decrease of end user p erformance while running in a dis tributed environment Vice versa relying on a distributedmemory paradigm ie messagepassing results in a decrease of end user p erformance while running in a sharedmemory environment Thus in the attempt of developing scalable and p ortable parallel application software programmers are concerned with a tradeo The purp ose of the VOTE system is to supp ort application programmers confronted with the tradeo b etween sharedmemory and messagepassing com munication Sharedmemory communication is physically limited to op erate lo cally only Therefore to implement a global sharedmemory communication fa cility VOTE supp orts a highly ecient core system The core system implements a global address space by means of VSM At default a multiple readersingle writer invalidationbased sequential consistency maintenance is provided A ex ible set of functions is available allowing many p owerful optimizations to pro grams running under control of sequential consistency As a step b eyond VOTE supp orts messagepassing communication MPI standard through a functional enrichment of its VSM system Thus VOTE provides to the applica tion programmer architectural transparency by means of sequential consistency and a messagepassing communication system The rest of this pap er is organized as follows Section deals with the prin ciple problem of supp orting b oth communication paradigms Afterwards Sec tion covers an application case study by demonstrating the implementation of successive overrelaxation dynamically changing b etween the two communication paradigms Section discusses the achieved runtime results Related works are presented in Section and Section concludes the pap er Communication Paradigms To days trend in the design of scalable highp erformance computing systems is based on a co op eration b etween software and hardware to supp ort b oth an ecient sharedmemory and messagepassing abstraction The main problem of this co op eration is due to the incompatibility of the two views of consistency semantics as given by the two communication paradigms Messagepassing is an explicit communication style lacking any measures of implicit consistency maintenance Global eects in the sense of making changes to some logically shared but physically distributed state are achieved only when the resp ective parallel application program explicitly calls commu nication functions In contrast changing p ortions of the logically shared but physically lo cal state works implicit by simply readingwriting the memory cells ie program variables The sharedmemory communication paradigm is a single readersingle writer SRSW scheme o ccasionally extended by cachebased sharedmemory machines to a multiple readersingle writer scheme Every mo difying memory access works implicit However since for p erformance reasons data replication usually takes place this paradigm requires either invalidating or up dating the data copies held in the caches or the memories These measures of consistency maintenance im plicity inuence the op eration of all pro cesses and generally impact the runtime b ehavior of a sharedmemory parallel program A paradigm shift from sharedmemory to messagepassing communication is almost straightforward It makes public ie global shared data private ie lo cal and thus nonshared The pro cess shifting to the messagepassing paradigm will b e deleted from the directory information maintained by the VSM sys tem to keep track of p otential data copy holders Due to the paradigm shift programming complexity increases signicantly Running under control of the sharedmemory paradigm consistency was ensured by the VSM system In the messagepassing case the programmer b ecomes resp onsible maintaining a glob ally consistent view of the common data sets A paradigm shift from messagepassing to sharedmemory is not that straightforward The main problem is the unication of the lo cal data copies into a single data image In other words private data copies are made public or invalidated and the directory information is up dated In most cases this task cannot b e p erformed automatically b ecause of the lack of information ab out the order of mo dications done by the individual pro cesses In the current imple mentation of VOTE it is therefore up to the application program to announce the lo cality of consistent data Case Study Successive Overrelaxation This section presents two implementations of a successive overrelaxation algo rithm based on the red black technique This kind of application is a typical representative for parallel numerical algorithms op erating on dense data struc tures The amount of data b eing shared is little and increases only with the num b er of computing no des Data sharing is p erformed only by a pair of pro cesses that is each pro cess communicates with two other pro cesses the predecessor and successor Merely the rst and last pro cess communicate only with a single pro cess The following subsections present two implementations of a SPMDbased single programmultiple data successive overrelaxation program The rst co de communicates via sharedmemory and uses a global barrier synchronization The second co de takes full advantage of the communication supp ort system VOTE that is it changes the communication paradigm from sharedmemory to message passing and vice versa SharedMemory Communication The co de example in Fig uses sharedmemory communication and is a very compact implementation of a successive overrelaxation The red black tech nique is used to avoid overwriting the old value of a matrix element b efore it is used for the computation of the new values of the resp ective neighb or Therefore computation is done only on every other row p er iteration line This approach requires two instead of one global barrier synchronization lines void sor float rMN float bMN int f int t int i j k while converged rbft for j f j t j for k k N k bjk rjkrjkrjkrjk if j t break for k k N k bjk rjkrjkrjkrjk votebarrier for j f j t j for k k N k rjk bjkbjkbjkbjk if j t break for k k N k rjk bjkbjkbjkbjk votebarrier Fig Sharedmemory version and The algorithm continues executing the outer lo op line until the check of convergence evaluates to true line Variables f and t line implement a blo ck distribution of matrix r and b These variables are calculated on a p erpro cess basis and are a function of the size of a matrix the total numb er of pro cesses and the logical identication of a pro cess The computation steps in lines and lines pro duce data used as input for the pro cesses with adjacent values of the variables f and t When the barrier synchronization falls that data is consumed by the communication partner MessagePassing Communication The second implementation in Fig comprises three parts The rst part lines implements a switch from sharedmemory to messagepassing com munication This requires a barrier synchronization making sure that all pro cesses are ready for leaving consistency maintenance of VOTE Line compute Comm rank the working set of a pro cess using the logical identication MPI and the total numb er of computing pro cesses MPI Comm size Calling access is to ensure
Load more

Recommended publications
  • Memory Consistency
    Lecture 9: Memory Consistency Parallel Computing Stanford CS149, Winter 2019 Midterm ▪ Feb 12 ▪ Open notes ▪ Practice midterm Stanford CS149, Winter 2019 Shared Memory Behavior ▪ Intuition says loads should return latest value written - What is latest? - Coherence: only one memory location - Consistency: apparent ordering for all locations - Order in which memory operations performed by one thread become visible to other threads ▪ Affects - Programmability: how programmers reason about program behavior - Allowed behavior of multithreaded programs executing with shared memory - Performance: limits HW/SW optimizations that can be used - Reordering memory operations to hide latency Stanford CS149, Winter 2019 Today: what you should know ▪ Understand the motivation for relaxed consistency models ▪ Understand the implications of relaxing W→R ordering Stanford CS149, Winter 2019 Today: who should care ▪ Anyone who: - Wants to implement a synchronization library - Will ever work a job in kernel (or driver) development - Seeks to implement lock-free data structures * - Does any of the above on ARM processors ** * Topic of a later lecture ** For reasons to be described later Stanford CS149, Winter 2019 Memory coherence vs. memory consistency ▪ Memory coherence defines requirements for the observed Observed chronology of operations on address X behavior of reads and writes to the same memory location - All processors must agree on the order of reads/writes to X P0 write: 5 - In other words: it is possible to put all operations involving X on a timeline such P1 read (5) that the observations of all processors are consistent with that timeline P2 write: 10 ▪ Memory consistency defines the behavior of reads and writes to different locations (as observed by other processors) P2 write: 11 - Coherence only guarantees that writes to address X will eventually propagate to other processors P1 read (11) - Consistency deals with when writes to X propagate to other processors, relative to reads and writes to other addresses Stanford CS149, Winter 2019 Coherence vs.
    [Show full text]
  • Memory Consistency: in a Distributed Outline for Lecture 20 Memory System, References to Memory in Remote Processors Do Not Take Place I
    Memory consistency: In a distributed Outline for Lecture 20 memory system, references to memory in remote processors do not take place I. Memory consistency immediately. II. Consistency models not requiring synch. This raises a potential problem. Suppose operations that— Strict consistency Sequential consistency • The value of a particular memory PRAM & processor consist. word in processor 2’s local memory III. Consistency models is 0. not requiring synch. operations. • Then processor 1 writes the value 1 to that word of memory. Note that Weak consistency this is a remote write. Release consistency • Processor 2 then reads the word. But, being local, the read occurs quickly, and the value 0 is returned. What’s wrong with this? This situation can be diagrammed like this (the horizontal axis represents time): P1: W (x )1 P2: R (x )0 Depending upon how the program is written, it may or may not be able to tolerate a situation like this. But, in any case, the programmer must understand what can happen when memory is accessed in a DSM system. Consistency models: A consistency model is essentially a contract between the software and the memory. It says that iff they software agrees to obey certain rules, the memory promises to store and retrieve the expected values. © 1997 Edward F. Gehringer CSC/ECE 501 Lecture Notes Page 220 Strict consistency: The most obvious consistency model is strict consistency. Strict consistency: Any read to a memory location x returns the value stored by the most recent write operation to x. This definition implicitly assumes the existence of a global clock, so that the determination of “most recent” is unambiguous.
    [Show full text]
  • Towards Shared Memory Consistency Models for Gpus
    TOWARDS SHARED MEMORY CONSISTENCY MODELS FOR GPUS by Tyler Sorensen A thesis submitted to the faculty of The University of Utah in partial fulfillment of the requirements for the degree of Bachelor of Science School of Computing The University of Utah March 2013 FINAL READING APPROVAL I have read the thesis of Tyler Sorensen in its final form and have found that (1) its format, citations, and bibliographic style are consistent and acceptable; (2) its illustrative materials including figures, tables, and charts are in place. Date Ganesh Gopalakrishnan ABSTRACT With the widespread use of GPUs, it is important to ensure that programmers have a clear understanding of their shared memory consistency model i.e. what values can be read when issued concurrently with writes. While memory consistency has been studied for CPUs, GPUs present very different memory and concurrency systems and have not been well studied. We propose a collection of litmus tests that shed light on interesting visibility and ordering properties. These include classical shared memory consistency model properties, such as coherence and write atomicity, as well as GPU specific properties e.g. memory visibility differences between intra and inter block threads. The results of the litmus tests are determined by feedback from industry experts, the limited documentation available and properties common to all modern multi-core systems. Some of the behaviors remain unresolved. Using the results of the litmus tests, we establish a formal state transition model using intuitive data structures and operations. We implement our model in the Murphi modeling language and verify the initial litmus tests.
    [Show full text]
  • Consistency Models • Data-Centric Consistency Models • Client-Centric Consistency Models
    Consistency and Replication • Today: – Consistency models • Data-centric consistency models • Client-centric consistency models Computer Science CS677: Distributed OS Lecture 15, page 1 Why replicate? • Data replication: common technique in distributed systems • Reliability – If one replica is unavailable or crashes, use another – Protect against corrupted data • Performance – Scale with size of the distributed system (replicated web servers) – Scale in geographically distributed systems (web proxies) • Key issue: need to maintain consistency of replicated data – If one copy is modified, others become inconsistent Computer Science CS677: Distributed OS Lecture 15, page 2 Object Replication •Approach 1: application is responsible for replication – Application needs to handle consistency issues •Approach 2: system (middleware) handles replication – Consistency issues are handled by the middleware – Simplifies application development but makes object-specific solutions harder Computer Science CS677: Distributed OS Lecture 15, page 3 Replication and Scaling • Replication and caching used for system scalability • Multiple copies: – Improves performance by reducing access latency – But higher network overheads of maintaining consistency – Example: object is replicated N times • Read frequency R, write frequency W • If R<<W, high consistency overhead and wasted messages • Consistency maintenance is itself an issue – What semantics to provide? – Tight consistency requires globally synchronized clocks! • Solution: loosen consistency requirements
    [Show full text]
  • Challenges for the Message Passing Interface in the Petaflops Era
    Challenges for the Message Passing Interface in the Petaflops Era William D. Gropp Mathematics and Computer Science www.mcs.anl.gov/~gropp What this Talk is About The title talks about MPI – Because MPI is the dominant parallel programming model in computational science But the issue is really – What are the needs of the parallel software ecosystem? – How does MPI fit into that ecosystem? – What are the missing parts (not just from MPI)? – How can MPI adapt or be replaced in the parallel software ecosystem? – Short version of this talk: • The problem with MPI is not with what it has but with what it is missing Lets start with some history … Argonne National Laboratory 2 Quotes from “System Software and Tools for High Performance Computing Environments” (1993) “The strongest desire expressed by these users was simply to satisfy the urgent need to get applications codes running on parallel machines as quickly as possible” In a list of enabling technologies for mathematical software, “Parallel prefix for arbitrary user-defined associative operations should be supported. Conflicts between system and library (e.g., in message types) should be automatically avoided.” – Note that MPI-1 provided both Immediate Goals for Computing Environments: – Parallel computer support environment – Standards for same – Standard for parallel I/O – Standard for message passing on distributed memory machines “The single greatest hindrance to significant penetration of MPP technology in scientific computing is the absence of common programming interfaces across various parallel computing systems” Argonne National Laboratory 3 Quotes from “Enabling Technologies for Petaflops Computing” (1995): “The software for the current generation of 100 GF machines is not adequate to be scaled to a TF…” “The Petaflops computer is achievable at reasonable cost with technology available in about 20 years [2014].” – (estimated clock speed in 2004 — 700MHz)* “Software technology for MPP’s must evolve new ways to design software that is portable across a wide variety of computer architectures.
    [Show full text]
  • An Overview on Cyclops-64 Architecture - a Status Report on the Programming Model and Software Infrastructure
    An Overview on Cyclops-64 Architecture - A Status Report on the Programming Model and Software Infrastructure Guang R. Gao Endowed Distinguished Professor Electrical & Computer Engineering University of Delaware [email protected] 2007/6/14 SOS11-06-2007.ppt 1 Outline • Introduction • Multi-Core Chip Technology • IBM Cyclops-64 Architecture/Software • Cyclops-64 Programming Model and System Software • Future Directions • Summary 2007/6/14 SOS11-06-2007.ppt 2 TIPs of compute power operating on Tera-bytes of data Transistor Growth in the near future Source: Keynote talk in CGO & PPoPP 03/14/07 by Jesse Fang from Intel 2007/6/14 SOS11-06-2007.ppt 3 Outline • Introduction • Multi-Core Chip Technology • IBM Cyclops-64 Architecture/Software • Programming/Compiling for Cyclops-64 • Looking Beyond Cyclops-64 • Summary 2007/6/14 SOS11-06-2007.ppt 4 Two Types of Multi-Core Architecture Trends • Type I: Glue “heavy cores” together with minor changes • Type II: Explore the parallel architecture design space and searching for most suitable chip architecture models. 2007/6/14 SOS11-06-2007.ppt 5 Multi-Core Type II • New factors to be considered –Flops are cheap! –Memory per core is small –Cache-coherence is expensive! –On-chip bandwidth can be enormous! –Examples: Cyclops-64, and others 2007/6/14 SOS11-06-2007.ppt 6 Flops are Cheap! An example to illustrate design tradeoffs: • If fed from small, local register files: 64-bit FP – 3200 GB/s, 10 pJ/op unit – < $1/Gflop (60 mW/Gflop) (drawn to a 64-bit FPU is < 1mm^2 scale) and ~= 50pJ • If fed from global on-chip memory: Can fit over 200 on a chip.
    [Show full text]
  • Constraint Graph Analysis of Multithreaded Programs
    Journal of Instruction-Level Parallelism 6 (2004) 1-23 Submitted 10/03; published 4/04 Constraint Graph Analysis of Multithreaded Programs Harold W. Cain [email protected] Computer Sciences Dept., University of Wisconsin 1210 W. Dayton St., Madison, WI 53706 Mikko H. Lipasti [email protected] Dept. of Electrical and Computer Engineering, University of Wisconsin 1415 Engineering Dr., Madison, WI 53706 Ravi Nair [email protected] IBM T.J. Watson Research Laboratory P.O. Box 218, Yorktown Heights, NY 10598 Abstract This paper presents a framework for analyzing the performance of multithreaded programs using a model called a constraint graph. We review previous constraint graph definitions for sequentially consistent systems, and extend these definitions for use in analyzing other memory consistency models. Using this framework, we present two constraint graph analysis case studies using several commercial and scientific workloads running on a full system simulator. The first case study illus- trates how a constraint graph can be used to determine the necessary conditions for implementing a memory consistency model, rather than conservative sufficient conditions. Using this method, we classify coherence misses as either required or unnecessary. We determine that on average over 30% of all load instructions that suffer cache misses due to coherence activity are unnecessarily stalled because the original copy of the cache line could have been used without violating the mem- ory consistency model. Based on this observation, we present a novel delayed consistency imple- mentation that uses stale cache lines when possible. The second case study demonstrates the effects of memory consistency constraints on the fundamental limits of instruction level parallelism, com- pared to previous estimates that did not include multiprocessor constraints.
    [Show full text]
  • Memory Consistency Models
    Memory Consistency Models David Mosberger TR 93/11 Abstract This paper discusses memory consistency models and their in¯uence on software in the context of parallel machines. In the ®rst part we review previous work on memory consistency models. The second part discusses the issues that arise due to weakening memory consistency. We are especially interested in the in¯uence that weakened consistency models have on language, compiler, and runtime system design. We conclude that tighter interaction between those parts and the memory system might improve performance considerably. Department of Computer Science The University of Arizona Tucson, AZ 85721 ¡ This is an updated version of [Mos93] 1 Introduction increases. Shared memory can be implemented at the hardware Traditionally, memory consistency models were of in- or software level. In the latter case it is usually called terest only to computer architects designing parallel ma- DistributedShared Memory (DSM). At both levels work chines. The goal was to present a model as close as has been done to reap the bene®ts of weaker models. We possible to the model exhibited by sequential machines. conjecture that in the near future most parallel machines The model of choice was sequential consistency (SC). will be based on consistency models signi®cantly weaker ¡ Sequential consistency guarantees that the result of any than SC [LLG 92, Sit92, BZ91, CBZ91, KCZ92]. execution of processors is the same as if the opera- The rest of this paper is organized as follows. In tions of all processors were executed in some sequential section 2 we discuss issues characteristic to memory order, and the operations of each individual processor consistency models.
    [Show full text]
  • The Openmp Memory Model
    The OpenMP Memory Model Jay P. Hoeflinger1 and Bronis R. de Supinski2 1 Intel, 1906 Fox Drive, Champaign, IL 61820 [email protected] http://www.intel.com 2 Lawrence Livermore National Laboratory, P.O. Box 808, L-560, Livermore, California 94551-0808* [email protected] http://www.llnl.gov/casc Abstract. The memory model of OpenMP has been widely misunder- stood since the first OpenMP specification was published in 1997 (For- tran 1.0). The proposed OpenMP specification (version 2.5) includes a memory model section to address this issue. This section unifies and clarifies the text about the use of memory in all previous specifications, and relates the model to well-known memory consistency semantics. In this paper, we discuss the memory model and show its implications for future distributed shared memory implementations of OpenMP. 1 Introduction Prior to the OpenMP version 2.5 specification, no separate OpenMP Memory Model section existed in any OpenMP specification. Previous specifications had scattered information about how memory behaves and how it is structured in an OpenMP pro- gram in several sections: the parallel directive section, the flush directive section, and the data sharing attributes section, to name a few. This has led to misunderstandings about how memory works in an OpenMP program, and how to use it. The most problematic directive for users is probably the flush directive. New OpenMP users may wonder why it is needed, under what circumstances it must be used, and how to use it correctly. Perhaps worse, the use of explicit flushes often confuses even experienced OpenMP programmers.
    [Show full text]
  • Optimization of Memory Management on Distributed Machine Viet Hai Ha
    Optimization of memory management on distributed machine Viet Hai Ha To cite this version: Viet Hai Ha. Optimization of memory management on distributed machine. Other [cs.OH]. Institut National des Télécommunications, 2012. English. NNT : 2012TELE0042. tel-00814630 HAL Id: tel-00814630 https://tel.archives-ouvertes.fr/tel-00814630 Submitted on 17 Apr 2013 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. ABCDBEFAD EBDDDFADD ABCDEF EEEFAABCDEFCAADDABCDDA FCF EFECFF !" #$!"B ! "!#$!%" EC%&EF'%(' C*F+EEA ADDCDD A F !"DDDD#$% AC!AD&F'F(CA ))DC !*ADD+, F DDCADAD AD. !*(/AD !F DDCADA0, 1ADC !*2D3 C !,FF DDCDAADDD!AD4CAD !*4A)D4 DDCDAA.5 !D*(D+,!(( !"DDDD#$%4! 2012TELE0042 ABCD ABACDEAFFEBBADDEABFAEABFBFDAEFE EEABEFACAABAFAFFFFBAEABBA ADEFBFAEAABEBAADAEE EBABAFDB !"EABAFBABAEA#AEF $%&EBAEADFEFBEBADEAEAF AADA#BABBA!ABAFAAEABEBAFEFDBFB#E FB$%&EBAFDDEBBAEBF ABBEAFB$%&EBAEFDDEBBAFEA'F ACDAEABF (AADBBAFBBAEAABFDDEBA DEABFFFBAEF#EBAEFABEA ABABFBAFFBBAA#AEF ABCDEEFABCABBBFADFEFD DFEFBDDBAFFBFBDDBF DABBBBDAFEBBDEA ACA DEBDAAFB ABBB!CDCFBBEDCEFEDB " BB C E
    [Show full text]
  • Dag-Consistent Distributed Shared Memory
    Dag-Consistent Distributed Shared Memory Robert D. Blumofe Matteo Frigo Christopher F. Joerg Charles E. Leiserson Keith H. Randall MIT Laboratory for Computer Science 545 Technology Square Cambridge, MA 02139 Abstract relaxed models of shared-memory consistency have been developed [10, 12, 13] that compromise on semantics for We introduce dag consistency, a relaxed consistency model a faster implementation. By and large, all of these consis- for distributed shared memory which is suitable for multi- tency models have had one thingin common: they are ªpro- threaded programming. We have implemented dag consis- cessor centricº in the sense that they de®ne consistency in tency in software for the Cilk multithreaded runtime system terms of actions by physical processors. In this paper, we running on a Connection Machine CM5. Our implementa- introduce ªdagº consistency, a relaxed consistency model tion includes a dag-consistent distributed cactus stack for based on user-level threads which we have implemented for storage allocation. We provide empirical evidence of the Cilk[4], a C-based multithreadedlanguage and runtimesys- ¯exibility and ef®ciency of dag consistency for applications tem. that include blocked matrix multiplication, Strassen's ma- trix multiplication algorithm, and a Barnes-Hut code. Al- Dag consistency is de®ned on the dag of threads that make though Cilk schedules the executions of these programs dy- up a parallel computation. Intuitively, a read can ªseeº a namically, their performances are competitive with stati- write in the dag-consistency model only if there is some se- cally scheduled implementations in the literature. We also rial execution order consistentwiththedag in which theread sees the write.
    [Show full text]
  • Memory Models
    Chapter X Memory Models Leonidas I. Kontothanassis and Michael L. Scott Department of Computer Science University of Rochester Rochester, NY 14627-0226 fkthanasi,[email protected] May 1994 Contents 0.1 Memory Hardware Technology . 3 0.2 Memory System Architecture . 6 0.2.1 High-level Memory Architecture . 6 0.2.2 Basic Issues in Cache Design . 11 0.2.3 The Cache Coherence Problem . 12 0.3 User-Level Memory Models . 15 0.3.1 Shared Memory v. Message Passing . 15 0.3.2 Implementation of Shared-Memory Models . 17 0.3.3 Performance of Shared-Memory Models . 19 0.4 Memory Consistency Models . 22 0.5 Implementation and Performance of Memory Consistency Models . 27 0.5.1 Hardware Implementations . 27 0.5.2 Software Implementations . 30 0.6 Conclusions and Trends . 32 The rapid evolution of RISC microprocessors has left memory speeds lagging behind: as of early 1994, typical main memory (DRAM) access times are on the order of 100 processor 1 cycles, and this number is expected to increase for at least the next few years. Multiproces- sors, like uniprocessors, use memory to store both instructions and data. The cost of the typical memory access can therefore have a major impact on program execution time. To minimize the extent to which processors wait for memory, modern computer systems depend heavily on caches. Caches work extremely well for instructions: most programs spend most of their time in loops that fit entirely in the cache. As a result, instruction fetches usually have a negligible impact on a processor's average number of cycles per instruction (CPI).
    [Show full text]