Shared memory MIMD architecture

Introduction to MIMD Architectures:

Multiple direction watercourse, multiple informations watercourse ( MIMD ) machines have a figure of processors that function asynchronously and independently. At any clip, different processors may be put to deathing different instructions on different pieces of informations. MIMD architectures may be used in a figure of application countries such as computer-aided design/computer-aided fabrication, simulation, mold, and as communicating switches. MIMD machines can be of either shared memory or distributed memory classs. These categorizations are based on how MIMD processors entree memory. Shared memory machines may be of the bus-based, drawn-out, or hierarchal type. Distributed memory machines may hold hypercube or mesh interconnectedness strategies.


A type of multiprocessor architecture in which several direction rhythms may be active at any given clip, each independently taking instructions and operands into multiple treating units and runing on them in a coincident manner. Acronym for multiple-instruction-stream.multiple-data-stream.

Bottom of Form

( Multiple Instruction watercourse Multiple Data watercourse ) A computing machine that can treat two or more independent sets of instructions at the same time on two or more sets of informations. Computers with multiple CPUs or individual CPUs with double nucleuss are illustrations of MIMD architecture. Hyperthreading besides consequences in a certain grade of MIMD public presentation every bit good. Contrast with SIMD.

In calculating, MIMD ( Multiple Instruction watercourse, Multiple Data watercourse ) is a technique employed to accomplish correspondence. Machines utilizing MIMD have a figure of processors that function asynchronously and independently. At any clip, different processors may be put to deathing different instructions on different pieces of informations. MIMD architectures may be used in a figure of application countries such as computer-aided design/computer-aided fabrication, simulation, mold, and as communicating switches. MIMD machines can be of either shared memory or distributed memory classs. These categorizations are based on how MIMD processors entree memory. Shared memory machines may be of the bus-based, drawn-out, or hierarchal type. Distributed memory machines may hold hypercube or mesh interconnectedness strategies.

Multiple Instruction – Multiple Data

MIMD architectures have multiple processors that each execute an independent watercourse ( sequence ) of machine instructions. The processors execute these instructions by utilizing any accessible informations instead than being forced to run upon a individual, shared informations watercourse. Hence, at any given clip, an MIMD system can be utilizing as many different direction watercourses and informations watercourses as there are processors.

Although package processes put to deathing on MIMD architectures can be synchronized by go throughing informations among processors through an interconnectedness web, or by holding processors examine informations in a shared memory, the processors ‘ independent executing makes MIMD architectures asynchronous machines.

Shared Memory: Bus-based

MIMD machines with shared memory have processors which portion a common, cardinal memory. In the simplest signifier, all processors are attached to a coach which connects them to memory. This apparatus is called bus-based shared memory. Bus-based machines may hold another coach that enables them to pass on straight with one another. This extra coach is used for synchronism among the processors. When utilizing bus-based shared memory MIMD machines, merely a little figure of processors can be supported. There is contention among the processors for entree to shared memory, so these machines are limited for this ground. These machines may be incrementally expanded up to the point where there is excessively much contention on the coach.

Shared Memory: Extended

MIMD machines with extended shared memory effort to avoid or cut down the contention among processors for shared memory by subdividing the memory into a figure of independent memory units. These memory units are connected to the processsors by an interconnectedness web. The memory units are treated as a incorporate cardinal memory. One type of interconnectedness web for this type of architecture is a crossbar shift web. In this strategy, N processors are linked to M memory units which requires N times M switches. This is non an economically executable apparatus for linking a big figure of processors.

Shared Memory: Hierarchical

MIMD machines with hierarchal shared memory usage a hierarchy of coachs to give processors entree to each other ‘s memory. Processors on different boards may pass on through inter nodal coachs. Buss support communicating between boards. We use this type of architecture, the machine may back up over a 1000 processors.

In calculating, shared memory is memory that may be at the same time accessed by multiple plans with an purpose to supply communicating among them or avoid excess transcripts. Depending on context, plans may run on a individual processor or on multiple separate processors. Using memory for communicating inside a individual plan, for illustration among its multiple togss, is by and large non referred to as shared memory


In computing machine hardware, shared memory refers to a ( typically ) big block of random entree memory that can be accessed by several different cardinal treating units ( CPUs ) in a multiple-processor computing machine system.

A shared memory system is comparatively easy to plan since all processors portion a individual position of informations and the communicating between processors can be every bit fast as memory entrees to a same location.

The issue with shared memory systems is that many CPUs need fast entree to memory and will probably hoard memory, which has two complications:

  • CPU-to-memory connexion becomes a constriction. Shared memory computing machines can non scale really good. Most of them have ten or fewer processors.
  • Cache coherency: Whenever one cache is updated with information that may be used by other processors, the alteration needs to be reflected to the other processors, otherwise the different processors will be working with incoherent informations ( see cache coherency and memory coherency ) . Such coherency protocols can, when they work good, supply highly high-performance entree to shared information between multiple processors. On the other manus they can sometimes go overladen and go a constriction to public presentation.

The options to shared memory are distributed memory and distributed shared memory, each holding a similar set of issues. See besides Non-Uniform Memory Access.


In computing machine package, shared memory is either

  • A method of inter-process communicating ( IPC ) , i.e. a manner of interchanging informations between plans running at the same clip. One procedure will make an country in RAM which other procedures can entree, or
  • A method of conserving memory infinite by directing entrees to what would normally be transcripts of a piece of informations to a individual case alternatively, by utilizing practical memory functions or with expressed support of the plan in inquiry. This is most frequently used for shared libraries and for Execute in Place.

Shared Memory MIMD Architectures:

The distinguishing characteristic of shared memory systems is that no affair how many memory blocks are used in them and how these memory blocks are connected to the processors and address infinites of these memory blocks are unified into a planetary reference infinite which is wholly seeable to all processors of the shared memory system. Publishing a certain memory reference by any processor will entree the same memory block location. However, harmonizing to the physical organisation of the logically shared memory, two chief types of shared memory system could be distinguished:

Physically shared memory systems

Virtual ( or distributed ) shared memory systems

In physically shared memory systems all memory blocks can be accessed uniformly by all processors. In distributed shared memory systems the memory blocks are physically distributed among the processors as local memory units.

The three chief design issues in increasing the scalability of shared memory systems are:

  • Organization of memory
  • Design of interconnectedness webs
  • Design of cache coherent protocols

Cache Coherence:

Cache memories are introduced into computing machines in order to convey informations closer to the processor and hence to cut down memory latency. Caches widely accepted and employed in uniprocessor systems. However, in multiprocessor machines where several processors require a transcript of the same memory block.

The care of consistence among these transcripts raises the alleged cache coherency job which has three causes:

  • Sharing of writable informations
  • Procedure migration
  • I/O activity

From the point of position of cache coherency, informations constructions can be divided into three categories:

  • Read-only informations constructions which ne’er cause any cache coherency job. They can be replicated and placed in any figure of cache memory blocks without any job.
  • Shared writable informations constructions are the chief beginning of cache coherency jobs.
  • Private writable informations constructions pose cache coherency jobs merely in the instance of procedure migration.

There are several techniques to keep cache coherency for the critical instance, that is, shared writable informations constructions. The applied methods can be divided into two categories:

  • hardware-based protocols
  • software-based protocols

Software-based strategies normally introduce some limitations on the cachability of informations in order to forestall cache coherency jobs.

Hardware-based Protocols:

Hardware-based protocols provide general solutions to the jobs of cache coherency without any limitations on the cachability of informations. The monetary value of this attack is that shared memory systems must be extended with sophisticated hardware mechanisms to back up cache coherency. Hardware-based protocols can be classified harmonizing to their memory update policy, cache coherency policy, and interconnectedness strategy. Two types of memory update policy are applied in multiprocessors: write-through and write-back. Cache coherency policy is divided into write-update policy and write-invalidate policy.

Hardware-based protocols can be farther classified into three basic categories depending on the nature of the interconnectedness web applied in the shared memory system. If the web expeditiously supports broadcast medium, the alleged Snoopy cache protocol can be well exploited. This strategy is typically used in individual bus-based shared memory systems where consistence commands ( invalidate or update bids ) are broadcast via the coach and each cache ‘snoops ‘ on the coach for incoming consistence bids.

Large interconnectedness webs like multistage webs can non back up airing expeditiously and hence a mechanism is needed that can straight frontward consistence bids to those caches that contain a transcript of the updated information construction. For this intent a directory must be maintained for each block of the shared memory to administrate the existent location of blocks in the possible caches. This attack is called the directory strategy.

The 3rd attack attempts to avoid the application of the dearly-won directory strategy but still supply high scalability. It proposes multiple-bus webs with the application of hierarchal cache coherency protocols that are generalized or extended versions of the individual bus-based Snoopy cache protocol.

In depicting a cache coherency protocol the undermentioned definitions must be given:

  • Definition of possible provinces of blocks in caches, memories and directories.
  • Definition of bids to be performed at assorted read/write hit/miss actions.
  • Definition of province passages in caches, memories and directories harmonizing to the bids.
  • Definition of transmittal paths of bids among processors, caches, memories and directories.

Software-based Protocols:

Although hardware-based protocols offer the fastest mechanism for keeping cache consistence, they introduce a important excess hardware complexness, peculiarly in scalable multiprocessors. Software-based attacks represent a good and competitory via media since they require about negligible hardware support and they can take to the same little figure of annulment girls as the hardware-based protocols. All the software-based protocols rely on compiler aid.

The compiler analyses the plan and classifies the variables into four categories:

  • Read-only
  • Read-only for any figure of procedures and read-write for one procedure
  • Read-write for one procedure
  • Read-write for any figure of procedures.

Read-only variables can be cached without limitations. Type 2 variables can be cached merely for the processor where the read-write procedure tallies. Since merely one procedure uses type 3 variables it is sufficient to hoard them merely for that procedure. Type 4 variables must non be cached in software-based strategies. Variables demonstrate different behaviour in different plan subdivisions and hence the plan is normally divided into subdivisions by the compiler and the variables are categorized independently in each subdivision. More than that, the compiler generates instructions that control the cache or entree the cache explicitly based on the categorization of variables and codification cleavage. Typically, at the terminal of each plan subdivision the caches must be invalidated to guarantee that the variables are in a consistent province before get downing a new subdivision.

shared memory systems can be divided into four chief categories:

Uniform Memory Access ( UMA ) Machines:

Contemporary unvarying memory entree machines are small-size individual coach multiprocessors. Large UMA machines with 100s of processors and a shift web were typical in the early design of scalable shared memory systems. Celebrated representatives of that category of multiprocessors are the Denelcor HEP and the NYU Ultracomputer. They introduced many advanced characteristics in their design, some of which even today represent a important milepost in parallel computing machine architectures. However, these early systems do non incorporate either cache memory or local chief memory which turned out to be necessary to accomplish high public presentation in scalable shared memory systems

Non-Uniform Memory Access ( NUMA ) Machines:

Non-uniform memory entree ( NUMA ) machines were designed to avoid the memory entree constriction of UMA machines. The logically shared memory is physically distributed among the treating nodes of NUMA machines, taking to distributed shared memory architectures. On one manus these parallel computing machines became extremely scalable, but on the other manus they are really sensitive to data allotment in local memories. Accessing a local memory section of a node is much faster than accessing a distant memory section. Not by opportunity, the construction and design of these machines resemble in many ways that of distributed memory multicomputers. The chief difference is in the organisation of the address infinite. In multiprocessors, a planetary reference infinite is applied that is uniformly seeable from each processor ; that is, all processors can transparently entree all memory locations. In multicomputers, the reference infinite is replicated in the local memories of the processing elements. This difference in the address infinite of the memory is besides reflected at the package degree: distributed memory multicomputers are programmed on the footing of the message-passing paradigm, while NUMA machines are programmed on the footing of the planetary reference infinite ( shared memory ) rule.

The job of cache coherence does non look in distributed memory multicomputers since the message-passing paradigm explicitly handles different transcripts of the same information construction in the signifier of independent messages. In the shard memory paradigm, multiple entrees to the same planetary information construction are possible and can be accelerated if local transcripts of the planetary information construction are maintained in local caches. However, the hardware-supported cache consistence strategies are non introduced into the NUMA machines. These systems can hoard read-only codification and informations, every bit good as local informations, but non shared modifiable informations. This is the separating characteristic between NUMA and CC-NUMA multiprocessors. Consequently, NUMA machines are nearer to multicomputers than to other shared memory multiprocessors, while CC-NUMA machines look like existent shared memory systems.

In NUMA machines, like in multicomputers, the chief design issues are the organisation of processor nodes, the interconnectedness web, and the possible techniques to cut down distant memory entrees. Two illustrations of NUMA machines are the Hector and the Cray T3D multiprocessor.

  2. hypertext transfer protocol: // searchkeys=MIMD+architecture
  3. hypertext transfer protocol: //
  4. hypertext transfer protocol: //