Memory Consistency Memory Consistency Memory Consistency z Reads

Memory Consistency

Memory Consistency

Memory Consistency z Reads and writes of the shared memory face consistency problem z Need to achieve controlled consistency in memory events z Shared memory behavior determined by: y Program order y Memory access order z Challenges y Modern processors reorder operations y Compiler optimizations (scalar replacement, instruction rescheduling

Basic Concept z On a multiprocessor: y Concurrent instruction streams (threads) on different processors y Memory events performed by one process may create data to be used by another x. Events: read and write z Memory consistency model specifies how the memory events initiated by one process should be observed by other processes z Event ordering y Declare which memory access is allowed , which process should wait for a later access when processes compete

Uniprocessor vs. Multiprocessor Model

Understanding Program Order Initially X = 2 P 1 …. . r 0=Read(X) r 0=r 0+1 Write(r 0, X) …. . Possible execution sequences: P 1: r 0=Read(X) P 2: r 1=Read(X) P 1: r 0=r 0+1 P 1: Write(r 0, X) P 2: r 1=r 1+1 P 2: Write(r 1, X) x=3 P 2 …. . r 1=Read(x) r 1=r 1+1 Write(r 1, X) …… P 2: r 1=Read(X) P 2: r 1=r 1+1 P 2: Write(r 1, X) P 1: r 0=Read(X) P 1: r 0=r 0+1 P 1: Write(r 0, X) x=4

Interleaving P 1 a. A=1; b. Print B, C; P 2 c. B=1; d. Print A, C; P 3 e. C=1; f. Print A, B; switch A, B, C shared variables (initially 0) Shared Memory z Program orders of individual instruction streams may need to be modified because of interaction among them y. Finding optimum global memory order is an NP hard problem

Example P 1 a. A=1; b. Print B, C; P 2 c. B=1; d. Print A, C; P 3 e. C=1; f. Print A, B; switch A, B, C shared variables (initially 0) Shared Memory z Concatenate program orders in P 1, P 2 and P 3 y 6 -tuple binary strings (64 output combinations) y (a, b, c, d, e, f) => (001011) (in order execution) y (a, c, e, b, d, f) => (111111) (in order execution) y (b, d, f, e, a, c) => (000000) (out of order execution) x 6! (720 possible permutations)

Mutual exclusion problem zmutual exclusion problem in concurrent programming yallow two threads to share a single-use resource without conflict, using only shared memory for communication. yavoid the strict alternation of a naive turntaking algorithm

Definition z If two processes attempt to enter a critical section at the same time, allow only one process in, based on whose turn it is. z If one process is already in the critical section, the other process will wait for the first process to exit. z How would you implement this without ymutual exclusion, yfreedom from deadlock, and yfreedom from starvation.

Solution: Dekker’s Algorithm z. This is done by the use of two flags f 0 and f 1 which indicate an intention to enter the critical section and a turn variable which indicates who has priority between the two processes.

flag[0] : = false flag[1] : = false turn : = 0 // or 1 P 0 flag[0] : = true while flag[1] = true { if turn ≠ 0 { flag[0] : = false while turn ≠ 0 { } flag[0] : = true } } // critical section. . . turn : = 1 flag[0] : = false // remainder // section P 1 flag[1] : = true while flag[0] = true { if turn ≠ 1 { flag[1] : = false while turn ≠ 1 { } flag[1] : = true } } // critical section. . . turn : = 0 flag[1] : = false // remainder // section

Disadvantages zlimited to two processes zmakes use of busy waiting instead of process suspension. z. Modern CPUs execute their instructions in an out-of-order fashion, yeven memory accesses can be reordered

Peterson’s Algorithm flag[0] = 0; flag[1] = 0; turn; P 0 flag[0] = 1; turn = 1; while (flag[1] == 1 && turn == 1) { // busy wait } // critical section. . . // end of critical section flag[0] = 0; P 1 flag[1] = 1; turn = 0; while (flag[0] == 1 && turn == 0) { // busy wait } // critical section. . . // end of critical section flag[1] = 0;

Lamport's bakery algorithm z a bakery with a numbering machine ythe 'customers' will be threads, identified by the letter i, obtained from a global variable. ymore than one thread might get the same number // declaration and initial values of global variables Entering: array [1. . NUM_THREADS] of bool = {false}; Number: array [1. . NUM_THREADS] of integer = {0}; 1 lock(integer i) { 2 Entering[i] = true; 3 Number[i] = 1 + max(Number[1], . . . , Number[NUM_THREADS]); 4 Entering[i] = false; 5 for (j = 1; j <= NUM_THREADS; j++) { 6 // Wait until thread j receives its number: 7 while (Entering[j]) { /* nothing */ } 8 // Wait until all threads with smaller numbers or with the same 9 // number, but with higher priority, finish their work: 10 while ((Number[j] != 0) && ((Number[j], j) < (Number[i], i))) { 11 /* nothing */ 12 } 13 } 14 } 15 unlock(integer i) { 16 Number[i] = 0; 17 } 18 Thread(integer i) { 19 while (true) { 20 lock(i); 21 // The critical section goes here. . . 22 unlock(i); 23 // non-critical section. . . 24 } 25 }

Models Strict Consistency: Read always returns with most recent Write to same address Sequential Consistency: The result of any execution appears as the interleaving of individual programs strictly in sequential program order Processor Consistency: Writes issued by each processor are in program order, but writes from different processors can be out of order (Goodman) Weak Consistency: Programmer uses synch operations to enforce sequential consistency (Dubois) Reads from each processor is not restricted More opportunities for pipelining

Relationship to Cache Coherence Protocol z Cache coherence protocol must observe the constraints imposed by the memory consistency model y. Ex: Read hit in a cache x. Reading without waiting for the completion of a previous write my violate sequential consistency z Cache coherence protocol provides a mechanism to propagate the newly written value z Memory consistency model places an additional constraint on when the value can be propagated to a given processor

Latency Tolerance z. Scalable systems y. Distributed shared memory architecture y. Access to remote memory: long latency y. Processor speed vs. the memory and interconnect z. Need for y. Latency reduction, avoidance, hiding

Latency Avoidance z. Organize user applications at architectural, compiler or application levels to achieve program/data locality z. Possible when applications exhibit: y. Temporal or spatial locality z. How do you enhance locality?

Locality Enhancement z Architectural support: y. Cache coherency protocols, memory consistency models, fast message passing, etc. z User support y. High Performance Fortran: program instructs compiler how to allocate the data (example ? ) z Software support y. Compiler performs certain transformations x. Example?

Latency Reduction z. What if locality is limited? z. Data access is dynamically changing? y. For ex: sorting algorithms z. We need latency reduction mechanisms y. Target communication subsystem x. Interconnect x. Network interface x. Fast communication software • Cluster: TCP, UDP, etc

Latency Hiding z Hide communication latency within computation y. Overlapping techniques x. Prefetching techniques • Hide read latency x. Distributed coherent caches • Reduce cache misses • Shorten time to retrieve clean copy x. Multiple context processors • Switch from one context to another when long-latency operations is encountered (hardware supported multithreading)

Memory Delays z SMP y high in multiprocessors due to added contention for shared resources such as a shared bus and memory modules z Distributed y are even more pronounced in distributed-memory multiprocessors where memory requests may need to be satisfied across an interconnection network. z By masking some or all of these significant memory latencies, prefetching can be an effective means of speeding up multiprocessor applications

Data Prefetching z. Overlapping computation with memory accesses y. Rather than waiting for a cache miss to perform a memory fetch, data prefetching anticipates such misses and issues a fetch to the memory system in advance of the actual memory reference.

Cache Hierarchy z. Popular latency reducing technique z. But still common for scientific programs to spend more than half their run times stalled on memory requests ypartially a result of the “on demand” fetch policy xfetch data into the cache from main memory only after the processor has requested a word and found it absent from the cache.

Why do scientific applications exhibit poor cache utilization? z Is something wrong with the principle of locality? z The traversal of large data arrays is often at the heart of this problem. z Temporal locality in array computations y once an element has been used to compute a result, it is often not referenced again before it is displaced from the cache to make room for additional array elements. z Sequential array accesses patterns exhibit a high degree of spatial locality, many other types of array access patterns do not. y For example, in a language which stores matrices in row-major order, a row-wise traversal of a matrix will result in consecutively referenced elements being widely separated in memory. Such strided reference patterns result in low spatial locality if the stride is greater than the cache block size. In this case, only one word per cache block is actually used while the remainder of the block remains untouched even though cache space has been allocated for it.

Memory references r 1, r 2 and r 3 not in the cache Time: Computation and memory references satisfied within the cache hierarchy main memory access time

Challenges z Cache pollution y Data arrives early enough to hide all of the memory latency y Data must be held in the processor cache for some period of time before it is used by the processor. y During this time, the prefetched data are exposed to the cache replacement policy and may be evicted from the cache before use. y Moreover, the prefetched data may displace data in the cache that is currently in use by the processor. z Memory bandwidth y Back to figure: x No prefetch: the three memory requests occur within the first 31 time units of program startup, x With prefetch: these requests are compressed into a period of 19 time units. y By removing processor stall cycles, prefetching effectively increases the frequency of memory requests issued by the processor. y Memory systems must be designed to match this higher bandwidth to avoid becoming saturated and nullifying the benefits of prefetching.

Spatial Locality z. Block transfer is a way of prefetching (1960 s) z. Software prefetching later (1980 s)

Binding Prefetch z Non-blocking load instructions ythese instructions are issued in advance of the actual use to take advantage of the parallelism between the processor and memory subsystem. y. Rather than loading data into the cache, however, the specified word is placed directly into a processor register. z the value of the prefetched variable is bound to a named location at the time the prefetch is issued.

Software-Initiated Data Prefetching z Some form of fetch instruction ycan be as simple as a load into a processor register z Fetches are non-blocking memory operations y. Allow prefetches to bypass other outstanding memory operations in the cache. z Fetch instructions cannot cause exceptions z The hardware required to implement softwareinitiated prefetching is modest

Prefetch Challenges z prefetch scheduling. yjudicious placement of fetch instructions within the target application. ynot possible to precisely predict when to schedule a prefetch so that data arrives in the cache at the moment it will be requested by the processor yuncertainties not predictable at compile time xcareful consideration when statically scheduling prefetch instructions. ymay be added by the programmer or by the compiler during an optimization pass. xprogramming effort ?

Suitable spots for “Fetch” zmost often used within loops responsible for large array calculations. ycommon in scientific codes, yexhibit poor cache utilization ypredictable array referencing patterns.

Example: How to solve these two issues? software piplining assume a four-word cache block Issues: Cache misses during the first iteration Unnecessary prefetches in the last iteration of the unrolled loop

Assumptions z implicit assumption y Prefetching one iteration ahead of the data’s actual use is sufficient to hide the latency z What if the loops contain small computational bodies. y Define prefetch distance xinitiate prefetches d iterations before the data is referenced x. How do you determine “d”? • Let – “l” be the average cache miss latency, measured in processor cycles, – “s” be the estimated cycle time of the shortest possible execution path through one loop iteration, including the prefetch overhead. yd

Revisiting the example zlet us assume an average miss latency of 100 processor cycles and a loop iteration time of 45 cycles yd=3 (handle a prefetch distance of three)

Case Study z Given a distributed-shared multiprocessor z let’s define a remote access cache (RAC) y Assume that RAC is located at the network interface of each node y Motivation: prefetched remote data could be accessed at a speed comparable to that of local memory while the processor cache hierarchy was reserved for demand-fetched data. z Which one is better: Having RAC or pretefetching data directly into the processor cache hierarchy? y Despite significantly increasing cache contention and y reducing overall cache space, y The latter approach results in higher cache hit rates, xdominant performance factor.

Case Study z Transfer of individual cache blocks across the interconnection network of a multiprocessor yields low network efficiency y what if we propose transferring prefetched data in larger units? z Method: a compiler schedules a single prefetch command before the loop is entered rather than software pipelining prefetches within a loop. y transfer of large blocks of remote memory used within the loop body y prefetched into local memory to prevent excessive cache pollution. z Issues: y binding prefetch since data stored in a processor’s local memory are not exposed to any coherency policy y imposes constraints on the use of prefetched data which, in turn, limits the amount of remote data that can be prefetched.

What about besides the “loops”? z Prefetching is normally restricted to loops y array accesses whose indices are linear functions of the loop indices y compiler must be able to predict memory access patterns when scheduling prefetches. y such loops are relatively common in scientific codes but far less so in general applications. z Irregular data structures y difficult to reliably predict when a particular data will be accessed y once a cache block has been accessed, there is less of a chance that several successive cache blocks will also be requested when data structures such as graphs and linked lists are used. y comparatively high temporal locality x result in high cache utilization thereby diminishing the benefit of prefetching.

What is the overhead of fetch instructions? z require extra execution cycles z fetch source addresses must be calculated and stored in the processor y to avoid recalculation for the matching load or store instruction. x How: • Register space x Problem: • compiler will have less register space to allocate to other active variables. • fetch instructions increase register pressure • It gets worse when – the prefetch distance is greater than one – multiple prefetch addresses z code expansion y may degrade instruction cache performance. z software-initiated prefetching is done statically y unable to detect when a prefetched block has been prematurely evicted and needs to be re-fetched.

Hardware-Initiated Data Prefetching z. Prefetching capabilities without the need for programmer or compiler intervention. z. No changes to existing executables yinstruction overhead completely eliminated. zcan take advantage of run-time information to potentially make prefetching more effective.

Cache Blocks z Typically: fetch data from main memory into the processor cache in units of cache blocks. y multiple word cache blocks are themselves a form of data prefetching. y large cache blocks x. Effective prefetching vs cache pollution. y What is the complication for SMPs with private caches xfalse sharing: when two or more processors wish to access different words within the same cache block and at least one of the accesses is a store. xcache coherence traffic is generated to ensure that the changes made to a block by a store operation are seen by all processors caching the block. • Unnecessary traffic • Increasing the cache block size increases the likelihood of such occurances z How do we take advantage of spatial locality without introducing some of the problems associated with large cache blocks?

Sequential prefetching zone block lookahead (OBL) approach yinitiates a prefetch for block b+1 when block b is accessed. z. How is it different from doubling the block size? yprefetched blocks are treated separately with regard to the cache replacement and coherency policies.

OBL: Case Study z Assume that a large block contains one word which is frequently referenced and several other words which are not in use. z Assume that an LRU replacement policy is used, z What is the implication? y the entire block will be retained even though only a portion of the block’s data is actually in use. z How do we solve? y Replace large block with two smaller blocks, xone of them could be evicted to make room for more active data. xuse of smaller cache blocks reduces the probability of false sharing

OBL implementations z Based on “what type of access to block b initiates the prefetch of b+1” y prefetch on miss x. Initiates a prefetch for block b+1 whenever an access for block b results in a cache miss. x. If b+1 is already cached, no memory access is initiated y tagged prefetch algorithms x. Associates a tag bit with every memory block. x. Use this bit to detect • when a block is demand-fetched or • when a prefetched block is referenced for the first time. x. Then, next sequential block is fetched. y Which one is better in terms of reducing miss rate? Prefetch on miss vs tagged prefetch?

Prefetch on miss vs tagged prefetch Accessing three contiguous blocks strictly sequential access pattern:

Shortcoming of the OBL zprefetch may not be initiated far enough in advance of the actual use to avoid a processor memory stall. y. A sequential access stream resulting from a tight loop, for example, may not allow sufficient time between the use of blocks b and b+1 to completely hide the memory latency.

How do you solve this shortcoming? z. Increase the number of blocks prefetched after a demand fetch from one to “d” y. As each prefetched block, b, is accessed for the first time, the cache is interrogated to check if blocks b+1, . . . b+d are present in the cache z. What if d=1? What kind of prefetching is this? y. Tagged

Another technique with d-prefetch z d prefetched blocks are brought into a FIFO stream buffer before being brought into the cache. y As each buffer entry is referenced, it is brought into the cache while the remaining blocks are moved up in the queue and a new block is prefetched into the tail position. y If a miss occurs in the cache and the desired block is also not found at the head of the stream buffer, the buffer is flushed. z Advantage: y prefetched data are not placed directly into the cache, y avoids cache pollution. z Disadvantage: y requires that prefetched blocks be accessed in a strictly sequential order to take advantage of the stream buffer.

Tradeoffs of d-prefetching? z Good: increasing the degree of prefetching yreduces miss rates in sections of code that show a high degree of spatial locality z Bad yadditional traffic and cache pollution are generated by sequential prefetching during program phases that show little spatial locality. z What if are able to vary the “d”

Adaptive sequential prefetching z d is matched to the degree of spatial locality exhibited by the program at a particular point in time. z a prefetch efficiency metric is periodically calculated z Prefetch efficiency y ratio of useful prefetches to total prefetches x a useful prefetch occurs whenever a prefetched block results in a cache hit. z d is initialized to one, y incremented whenever efficiency exceeds a predetermined upper threshold y decremented whenever the efficiency drops below a lower threshold y If d=0, no prefetching z Which one is better? adaptive or tagged prefetching? y Miss ratio vs Memory traffic and contention

Sequential prefetching summary z Does sequential prefetching require changes to existing executables? z What about the hardware complexity? z Which one offers both simplicity and performance? y TAGGED z Compared to software-initiated prefetching, what might be the problem? y tend to generate more unnecessary prefetches. y Non-sequential access patterns are not good x Ex: such as scalar references or array accesses with large strides, will result in unnecessary prefetch requests x do not exhibit the spatial locality upon which sequential prefetching is based. z To enable prefetching of strided and other irregular data access patterns, several more elaborate hardware prefetching techniques have been proposed.

Prefetching with arbitrary strides z. Reference Prediction Table State: initial, transient, steady

RPT Entries State Transition

Matrix Multiplication Assume that starting addresses a=10000 b=20000 c=30000, and 1 word cache block After the first iteration of inner loop

Matrix Multiplication After the second iteration of inner loop Hits/misses?

Matrix Multiplication After the third iteration b and c hits provided that a prefetch of distance one is enough

RPT Limitations z. Prefetch distance to one loop iteration y. Loop entrance : miss y. Loop exit: unnecessary prefetch z. How can we solve this? y. Use longer distance y. Prefetch address = effective address + (stride x distance ) ywith lookahead program counter (LA-PC)

Summary z. Prefetches ytimely, useful, and introduce little overhead. z. Reduce secondary effects in the memory system zstrategies are diverse and no single strategy provides optimal performance

Summary z. Prefetching schemes are diverse. z. To help categorize a particular approach it is useful to answer three basic questions concerning the prefetching mechanism: y 1) When are prefetches initiated, y 2) Where are prefetched data placed, y 3) What is the unit of prefetch?

Software vs Hardware Prefetching z Prefetch instructions actually increase the amount of work done by the processor. z Hardware-based prefetching techniques do not require the use of explicit fetch instructions. y hardware monitors the processor in an attempt to infer prefetching opportunities. y no instruction overhead y generates more unnecessary prefetches than software-initiated schemes. xneed to speculate on future memory accesses without the benefit of compile-time information • Cache pollution • Consume memory bandwidth

Conclusions z Prefetches can be initiated either by y explicit fetch operation within a program (software initiated) y logic that monitors the processor’s referencing pattern (hardwareinitiated). z Prefetches must be timely. y issued too early x chance that the prefetched data will displace other useful data or be displaced itself before use. y issued too late x may not arrive before the actual memory reference and introduce stalls z Prefetches must be precise. y The software approach issues prefetches only for data that is likely to be used y Hardware schemes tend to fetch more data unnecessarily.

Conclusions z The decision of where to place prefetched data in the memory hierarchy y higher level of the memory hierarchy to provide a performance benefit. z The majority of schemes y prefetched data in some type of cache memory. z Prefetched data in processor registers y binding and additional constraints must be imposed on the use of the data. z Finally, multiprocessor systems can introduce additional levels into the memory hierarchy which must be taken into consideration.

Conclusions z Data can be prefetched in units of single words, cache blocks or larger blocks of memory. ydetermined by the organization of the underlying cache and memory system. z Uniprocessors and SMPs y. Cache blocks appropriate z Distributed memory multiprocessor ylarger memory blocks xto amortize the cost of initiating a data transfer across an interconnection network