CSC/ECE 506 Spring 2012/11a hn

From Expertiza_Wiki
Revision as of 19:59, 15 April 2012 by Hramach (talk | contribs)
Jump to navigation Jump to search

Distributed Shared Memory

The Distributed Shared Memory (DSM) system is a combination of a Shared Memory system and a Distributed Memory systems. The Shared Memory system has a group of processors sharing global memory which allows any processor to access the memory location[1]. The Distributed memory systems have a cluster each (i.e each processor has its own memory location ) and use message passing to communicate between the clusters[2] The DSM on the other hand has several processors sharing the same address space i.e a location in memory will be the same physical address for all processors. Unlike the Shared Memory System they do not actually use a global memory which is accessible by all processors, instead it is a logical abstraction of a single address space for different memory locations which can be accessed by all processors.

Shared Memory System.
Illustration of Shared memory system [3].
Distributed Memory System.
Illustration of Distributed memory system [3].

Distributed Shared Memory System.
Illustration of Distributed Shared memory system [3].

From the above pictures it can be noticed that in Distributed Systems each processor as its own memory,but in case of Distributed Shared memory though each processor has its own memory the memory is the same i.e The main memory of a cluster of processors is made to look like a single memory with a single address space. The Distributed systems are scalable for a large number of processors but is difficult to program because of message passing. The Shared Memory on the other hand is less scalable but is easy to program as all the processors can access each others data and makes parallel programming easier. Therefore DSM uses a more scalable and less expensive model by supporting the abstraction of shared memory by using message passing. The DSM allows the programmer to share and use variables without having to worry about their management. Hence a processor can access a address space held by other processors main memory. It allows end-users to use the shared memory without knowing the message passing,the idea is to allow inter processor communication which is invisible to the user.

A DSM system in general has clusters connected to a interconnection network. A cluster can have a uni-processor or multi-processor system with local memory and cache to remove memory latency. The local memory of a cluster is entirely or partially mapped to the DSM global memory system. A directory is maintained which has information of location of each data block and the state of the block. Directory organization can be that of a linked list,double linked list or a full map storage,the organization of directory will affect the system scalabilty[5].

A DSM system provides the shared memory abstraction by creating a software which would be cost efficient as it would not change anything in the cluster. Some systems provide the abstraction in hardware which could cost extra but the performance of hardware based DSM is better than software based [4].

Why not use Bus based multiprocessing for the DSM?

The Distributed Shared Memory(DSM) architecture requires a shared memory space. Message passing using interconnect networks is the preferred mode of communication to maintain coherence in a DSM. This is because

- Physically, the DSM is designed to allow a large number of nodes to interact with each other. With a bus-based system, the bus become simply too long to connect various nodes of a DSM. In a bus based multiprocessor for a smaller set of nodes having a common memory, latency of communication is not discernable. Also because the nodes of a bus-based system use snoopy protocols, clock frequency synchronization is also critical. With larger number of nodes and a longer bus, maintaining clock frequency becomes a bigger challenge. - Whatever cache coherence protocol is used, a larger number of nodes leads to a larger number of transactions on the bus. The bus using a shared channel to transfer cache blocks/lines, an increase in bus traffic causes greater latency in communcation. - It is preferable to design a DSM on a non-uniform memory access(NUMA) based architecture, because that would allow the use of memory to be addressed globally, but placed locally to each node.

For the above mentioned reasons, a message passing model is used for communication between nodes. A message passing model augmented with a point to point connection between nodes is also preferable. The following section explains why.

Not Bus based, so where does the coherence happen?

As explored in the previous section, it is not preferable to use a monlithic bus for communication. Instead, short wires are used for point-to-point interconnections. As a result, these wires can afford to have greater frequency clocked. The bandwidth can also now afford to increase.

Distributed Memory System.
Efficiency of point-to-point interconnect for a DSM [3].

As illustrated in the figure, a point-to-point interconnection is coupled with a directory based protocol for a DSM. Since the memory is distributed, a directory can be maintained, which has information about which caches have a copy of memory blocks. Having a directory based protocol, with each node having information about the blocks maintatined by the caches, each time a read/write is done by a cache of a node, based on it being a hit or miss, allows a node to contact the directory system, and then send updates/signals to only those nodes in the system, thus avoiding traffic.


Hardware DSM

To achieve the hardware DSM we need special kind of network interfaces and cache coherence circuits. There are 3 prominent groups of hardware which are interesting

  • Cache Coherent Non Uniform Memory Architecture (CC-NUMA)
  • Cache only Memory Architecture(COMA)
  • Reflective Memory systems

CC-NUMA being the most important is discussed here.


CC-NUMA

A CC-NUMA architecture has its shared virtual address space distributed between various clusters local memories so that the local processor as well as other processors in the network can access with different access time as it is a non uniform memory architecture. "The address space of the architecture maps onto the local memories of each cluster in a NUMA fashion[5]". The cache in each cluster is used to replicate blocks of memory which are reserved in remote hosts.

CC-NUMA.
Cache Coherent Non Uniform Memory Architecture [5].


There if a fixed physical memory location for each data block in CC-NUMA. Directory based architecture is used in which each cluster has equal parts of systems shared memory which is called the home property.The various coherence protocols, Memory consistency models like weak consistency and processor consistency are implemented in CC-NUMA. The CC-NUMA is simple and cost effective but when multiple processors want to access the same block in succession there can be a huge performance downfall for CC-NUMA [6].


The hardware DSM fairs better than software DSM in terms of performance but the cost of hardware is very high and hence changes are being made in the implementation of software DSM for better performance and hence this article will deal in detail with software DSM ,their performance and performance improvement measures.


Software components of a DSM

The software components of a DSM form a critical part of the DSM application. Robust Software design decisions better help the DSM system to work efficiently. The idea of a good software component for a DSM depends on factors like portability to different operating systems, efficient memory allocation, and good coherence protocols to support data up to date. As we shall see further, in modern day DSMs the software needs to co-operate with the underlying hardware in order to ensure ideal system performance.

In case of a distributed shared memory, it is imperative that a software component that can ease the burden of the programmer writing message passing programs. The software distributed shared memory is responsible for actually providing the abstraction of a shared memory over a hardware.

Fine- grained v Coarse grained approaches to software shared memory

In the early days of the DSM development, a so-called "fine-grained" approach was implemented which used virtual memory based pages. Page faults caused invalidations and updates to occur. But the challenges caused by this model are the fact that synchronization within a page causes great overhead.

The coarse-grained approach is more a VM-based approach while the fine-grained approach is mostly an instrumentation-based approach. This section looks at how fine-grained and coarse grained approaches perform under similar underlying hardware. A good example of a fine-grained approach is the Shasta approach, while an example for the coarse-grained approach is the Cashmere approach.

Shasta protocol

It is shown that the Shasta performs better in applications that involve synchronization at a finer level instruction-set, while Cashmere, predictably, has a better performance on applications that have a coarse level synchronization structure. At granularity levels larger than a cache line, Shasta performs better because Shasta has more useful data in the form of pages owing to its finer granularity. If large amounts of private data needs to be synchronized, then the coarser Cashmere is preferred. In case of false sharing, Cashmere, which allows aggregation of protocol operations, performs better. False sharing is defined as misses on memory/caches that occur due to the fact that multiple threads share different blocks on different words residing in the same cache block. But in case of applications that have a high level of false write-write sharing, page based finer protocols perform better, because the software overhead required to maintain coherence is removed. This is because coherence within nodes is managed by the hardware itself.

Shasta uses an interesting approach for inter and intra-node communication. Firstly, it uses timely checks to determine if misses occur and processes them in software. The shared address space is divided into blocks, which are different for different systems. The blocks are further divided into lines, and each line maintains its state information in a state table. These periodic checks for misses take up a mere 7 instructions. Stack and memory local to a thread is not checked because they are not shared among nodes. Optimizing these checks are done by aggregating checks for nearby blocks.

CC-NUMA.
The 7 instruction Periodic Miss check done in software[7].

The code first checks if the target address is in shared memory, and if it is not, goes ahead with executing the rest of the code. Otherwise, as shown it checks the state table entry for the location of the block. Optimization done to arrive at this code are to not save or restore registers and look for unused registers to perform the operation. Stack pointers and global pointers, which correspond to data private to a thread are not checked.

There are challenges that are posed with the above approach. Sometimes, these checks for misses create a huge overhead, often exceeding the time taken for sequential execution itself!

Optimization for improving performance of Shasta

There is an efficient optimization technique for communication between and within nodes. Within a node, race conditions are avoided by locking blocks individually during coherence protocol operations. Instead of a bus based system within a node, messages are sent between nodes instead. Processes must also wait until all operations on the node pending are complete. If the program already has provisions for handling race conditions, this time wait required for release can be removed. Messages from other processors are done through polling. No interrupt based mechanism is used for message handling because of a high cost associated with it. This way, the processor is free to do other tasks. Polling is done when waiting for messages, and the polling location is common to all processors within a node, which is very efficient. Also just three instructions are needed to do the polling activity.[7]

Cashmere protocol

Cashmere is purely a page-based protocol for a distributed shared memory system. It requires the application programmer to handle data race conditions. Shared memory accesses are handled using a lock mechanism of 'Acquire' and 'Release'. As the name suggests, 'Acquire' gains access to shared data, while 'Release' grants access to data it has previously been holding. Invalidations are sent during the release operation, and its effects are only seen during the next acquire, during which any other operations that depend on that update can proceed with acquiring the lock. Hence, Cashmere forces the programmer to take into account these factors.

Cashmere uses the Memory Channel(MC) network to maintain directories that possess information of its constituent pages. Page faults are used that allow requests to be made for an up to date copy of the block. Modifications made by each node, as a result of writes, are also kept a track of by the directory. A per-processor dirty list is maintained that has all the updates made by a processor to a particular page. If the list of writers does not include the home node, ownership of the home to the processor currently writing it. And if that processor is the only one holding a copy of that block, it is moved to an exclusive state.

Optimization for improving performance of Cashmere

Cashmere uses a concept of twin to keep track of local updates to a copy of the block. The twin page is created when the processor that is the home node does not hold the exclusive copy of the block. When a release is done, the twin is compared to the newly written node. The difference is what is present in the home node, which now holds the clean copy of the block . The processor holding the release sends write updates to all dirty sharers of the block. The releaser downgrades permission for writes maintained by all copies holding the dirty version of the block. Then, the list of all nodes who performed a write on the block is also cleared.

Cashmere uses hardware support for performing its coherence. The concept of coalescing, where updates from different processors on the same node are aggregated and updated, allows much reduced overhead instead of processing each write request one by one. The concept of diffs on these twin pages on the coarse grained blocks that Cashmere operates on is what allows this protocol to operate efficiently.

Challenges faced by Cashmere and Shasta

Cashmere makes heavy use of the Memory channel for communication between nodes. It depends on the programmer for maintaining causality of messages after a processor releases the lock. Thus Cashmere assumes global ordering of messages in order to work correctly. It is too expensive to move away from the MC protocol for the Cashmere, since protocol changes need to be made at the code level. Shasta on the other hand is designed for networks that support fast message passing between nodes.

Performance of DSM

Memory Consistency Model

A memory consistency model in a shared memory space gives the order in which the memory can be accessed. The model of memory consistency model chosen will affect the performance of the system. In DSM using a strict model make it easier for the programmer an example of strict memory consistency model is sequential consistency. When using sequential consistency the results are similar to executing all the instructions in sequential order and the instructions should be executed in the program order specified for each processor. These properties of sequential consistency decrease the performance of parallel programs and can cause an overhead for performance for programs which have a higher parallelism. The higher granularity leads to false sharing which is also another performance issue that can cause an overhead for software DSM.

Improvement for Memory Consistency Using a memory consistency model cannot be avoided and hence many relaxed consistency models were developed with a goal to improve performance of software DSM systems. Release consistency showed better performance than most other relaxed consistency models. Number of messages sent is also affecting the performance of software DSM as sending a message is more costly than hardware DSM. In a lazy implementation of Release consistency all the messages are buffered until a release occurs and are sent out as a single message which decreases the communication overhead also. The modifications done by a processor are also transmitted to only the processor which acquires the block and hence also decrease the amount of data sent along with number of messages sent.


Cache Coherence Protocol

In a DSM system to make sure that when changes are made in a page which is cached in multiple locations the change has to be updated to other copies or the other copies should be invalidated. The protocols are called write update or write invalidate based on the protocol followed to maintain a coherent view of shared data.

Write Invalidate vs Write update: Write invalidate protocol invalidates all the shared copies of a block when it is being modified. The disadvantage of write invalidate protocol is that it causes a lot of false sharing (even if different part of the block is modified all copies are invalidated) whether or not the modified block is shared by other processors.

In case if update protocols all the shared part of the block are updated immediately after an update, it reduces false sharing but it updates the shared blocks even if they are not in use.

Therefore the type of protocol should be chosen based on the application i.e., for applications with sequential sharing we can use write invalidate and write update can be sued for applications with tight sharing.

Snoopy vs Directory based: Traditionally for DSM snoopy and directory based are the two protocols used. Snoopy protocol provides scalability to point to point connections and is less scalable in bus based systems but it needs hardware support and is not widely used for software DSM. For data that is frequently read and written there is lot more traffic saving in using directory based protocol that snoopy. [8]

Organization of directory based protocol: In directory based protocols the performance of DSM mainly depends on the organization of the directory. There are two kinds of directory based protocols memory based directory and cache based directory. In memory based directory the directory is placed in the main memory. In cache based directory each cache has a part of the directory. Cache based directory is scalable but very complicated to implement. A centralized memory based directory will a cause bottle neck as only one node has the directory and all if many nodes need the information at the same time. The centralized directory is not scalable as well as the directory is away from most of the processors and this can also affect performance of the system as memory latency will increase with the increase in distance from the processor. [8]

Physical location of the directory: If main memory is used for storing the directories it causes contention of main memory bandwidth as directory information should be accessed along with data from the main memory. The type of storage used also is a concern for performance as DRAM is dense but has slower access and SRAM is less dense and is faster and hence based on the need of the application a different physical location is chosen.[8]


Additional memory to maintain the directory is still a performance issue for directory based protocols. As the number of nodes increase the amount of memory for directory increases. The scalability of the protocol is limited due to the complexity to keep the copies coherent and also the memory considerations.

Performance Improvement for Cache Coherence protocols

We first discuss how the directory based protocol can be improved and then provide another protocol which is not largely used but has a better performance than a directory based protocol

Directory based protocol:

Organization: A distributed directory approach can be used to remove the bottle neck and improve the performance of DSM over centralized directory. The distributed approach maintains the directory information on many nodes and makes sure no bottle necks occur and also by keeping the directory information of a block in its home node. This reduces the memory latency as we can easily find the home node of the block and hence easily identify the directory information for that block.[8]

Physical location: There is no proper improvement for this but a separate memory location can be maintained for directory information and can use separate wires for accessing information which will not occupy the bandwidth needed for data. But the disadvantage of this method is excess cost.[8]

Lock based protocol: A lock based cache coherence protocol is designed to eliminate the performance concerns caused by directory based systems in the DSM. In the directory based system the owner of the block had the responsibility of propagating the changes caused to all other processors sharing the block. In order to avoid broadcasting the changes to all the processors a directory had to be maintained which would cause memory overhead. To avoid this problem in the new protocol we have the associated processors for a block to actively fetch the modified block and this removes the need for having a directory as well as removes the overhead of invalidation and update messages. The processors that share a block should know when should they acquire a new value and it can be done using the Release consistency model which is the best memory consistency model. Hence the lock based protocol is more efficient and scalable than the directory based protocol extensively used in DSM’s.

Performance Improvement

References

  1. Shared Memory System
  2. Distributed Memory system
  3. Distributed Shared Memory Systems
  4. Providing Hardware DSMPerformance at Software DSM Cost
  5. Distributed Shared Memory: Concepts and Systems
  6. CC-NUMA
  7. Comparative Evaluation of Fine- and Coarse-Grain Software Distributed Shared Memory
  8. Fundamentals of Parallel Computer Architecture