Chapter - 5

Multiprocessors and Thread-Level Parallelism

We have seen the renewed interest in developing multiprocessors in early 2000:

- The slowdown in uniprocessor performance due to the diminishing returns in exploring
instruction-level parallelism.
- Difficulty to dissipate the heat generated by uniprocessors with high clock rates.
- Demand for high-performance servers where thread-level parallelism is natural.
For all these reasons multiprocessor architectures has become increasingly attractive.

A Taxonomy of Parallel Architectures

The idea of using multiple processors both to increase performance and to improve
availability dates back to the earliest electronic computers. About 30 years ago, Flynn
proposed a simple model of categorizing all computers that is still useful today. He
looked at the parallelism in the instruction and data streams called for by the instructions
at the most constrained component of the multiprocessor, and placed all computers in one
of four categories:
1.Single instruction stream, single data stream
(SISD)—This category is the uniprocessor.

PU – Processing Unit

Uniprocessors

2.Single instruction stream, multiple data streams

(SIMD)—The same instruction is executed by multiple processors using different data
streams. Each processor has its own data memory (hence multiple data), but there is a
single instruction memory and control processor, which fetches and dispatches
instructions. Vector architectures are the largest class of processors of this type.

3.Multiple instruction streams, single data stream (MISD)—No commercial

multiprocessor of this type has been built to date, but may be in the future. Some special
purpose stream processors approximate a limited form of this (there is only a single data
stream that is operated on by successive functional units).
4. Multiple instruction streams, multiple data streams (MIMD)—Each processor fetches
its own instructions and operates on its own data. The processors are often off-the-shelf
microprocessors. This is a coarse model, as some multiprocessors are hybrids of these
categories. Nonetheless, it is useful to put a framework on the design space.

1. MIMDs offer flexibility. With the correct hardware and software support, MIMDs
can function as single-user multiprocessors focusing on high performance for one
application, as multiprogrammed multiprocessors running many tasks simultaneously, or
as some combination of these functions.
2. MIMDs can build on the cost/performance advantages of off-the-shelf
microprocessors. In fact, nearly all multiprocessors built today use the same
microprocessors found in workstations and single-processor servers.
With an MIMD, each processor is executing its own instruction stream. In many cases,
each processor executes a different process. Recall from the last chapter, that a process is
an segment of code that may be run independently, and that the state of the process
contains all the information necessary to execute that program on a processor. In a
multiprogrammed environment, where the processors may be running independent tasks,
each process is typically independent of the processes on other processors.
It is also useful to be able to have multiple processors executing a single program and
sharing the code and most of their address space. When multiple processes share code
and data in this way, they are often called threads
. Today, the term thread is often used in a casual way to refer to multiple loci of
execution that may run on different processors, even when they do not share an address
space. To take advantage of an MIMD multiprocessor with n processors, we must
usually have at least n threads or processes to execute. The independent threads are
typically identified by the programmer or created by the compiler. Since the parallelism
in this situation is contained in the threads, it is called thread-level parallelism.
Threads may vary from large-scale, independent processes–for example, independent
programs running in a multiprogrammed fashion on different processors– to parallel
iterations of a loop, automatically generated by a compiler and each executing for
perhaps less than a thousand instructions. Although the size of a thread is important in
considering how to exploit thread-level parallelism efficiently, the important qualitative
distinction is that such parallelism is identified at a high-level by the software system and
that the threads consist of hundreds to millions of instructions that may be executed in
parallel. In contrast, instruction level parallelism is identified by primarily by the
hardware, though with software help in some cases, and is found and exploited one
instruction at a time.
Existing MIMD multiprocessors fall into two classes, depending on the number of
processors involved, which in turn dictate a memory organization and interconnect
strategy. We refer to the multiprocessors by their memory organization, because what
constitutes a small or large number of processors is likely to change over time.
The first group, which we call
centralized shared-memory architectures

Centralized shared memory architectures have at most a few dozen processors in 2000.
For multiprocessors with small processor counts, it is possible for the processors to share
a single centralized memory and to interconnect the processors and memory by a bus.
With large caches, the bus and the single memory, possibly with multiple banks, can
satisfy the memory demands of a small number of processors. By replacing a single bus
with multiple buses, or even a switch, a centralized shared memory design can be scaled
to a few dozen processors. Although scaling beyond that is technically conceivable,
sharing a centralized memory, even organized as multiple banks, becomes less attractive
as the number of processors sharing it increases.
Because there is a single main memory that has a symmetric relationship to all processors
and a uniform access time from any processor, these multiprocessors are often called
symmetric (shared-memory) multiprocessors ( SMPs), and this style of architecture is
sometimes called UMA for uniform memory access. This type of centralized shared-
memory architecture is currently by far the most popular organization.
The second group consists of multiprocessors with physically distributed memory. To
support larger processor counts, memory must be distributed among the processors rather
than centralized; otherwise the memory system would not be able to support the
bandwidth demands of a larger number of processors without incurring excessively long
access latency. With the rapid increase in processor performance and the associated
increase in a processor’s memory bandwidth requirements, the scale of multiprocessor for
which distributed memory is preferred over a single, centralized memory continues to
decrease in number (which is another reason not to use small and large scale). Of course,
the larger number of processors raises the need for a high bandwidth interconnect.

Distributed-memory multiprocessor
Distributing the memory among the nodes has two major benefits. First, it is a cost-
effective way to scale the memory bandwidth, if most of the accesses are to the local
memory in the node. Second, it reduces the latency for accesses to the local memory.
These two advantages make distributed memory attractive at smaller processor counts as
processors get ever faster and require more memory bandwidth and lower memory
latency. The key disadvantage for a distributed memory architecture is that
communicating data between processors becomes somewhat more complex and has
higher latency, at least when there is no contention, because the processors no longer
share a single centralized memory. As we will see shortly, the use of distributed memory
leads to two different paradigms for interprocessor communication.
Typically, I/O as well as memory is distributed among the nodes of the multiprocessor,
and the nodes may be small SMPs (2–8 processors). Although the use of multiple
processors in a node together with a memory and a network interface is quite useful from
the cost-efficiency viewpoint.
Challenges for Parallel Processing
• Limited parallelism available in programs
– Need new algorithms that can have better parallel performance
 • Suppose you want to achieve a speedup of 80 with 100 processors. What fraction
of the original computation can be sequential?

1
80 =
FractionParallel
+ (1 − FractionParallel )
100
FractionParallel = 0.9975

Data Communication Models for Multiprocessors

– shared memory: access shared address space implicitly via load and store
operations.
– message-passing: done by explicitly passing messages among the
processors
• can invoke software with Remote Procedure Call (RPC)

• often via library, such as MPI: Message Passing Interface

• also called "Synchronous communication" since communication

causes synchronization between 2 processes
Message-Passing Multiprocessor
- The address space can consist of multiple private address spaces that are
logically disjoint and cannot be addressed by a remote processor

- The same physical address on two different processors refers to two

different locations in two different memories.
Multicomputer (cluster):
- can even consist of completely separate computers connected on a LAN.
- cost-effective for applications that require little or no communication.
Symmetric Shared-Memory Architectures
Multilevel caches can substantially reduce the memory bandwidth demands of a
processor.
This is extremely
- Cost-effective
- This can work as plug in play by placing the processor and cache sub-system on
a board into the bus backplane.

Developed by
• IBM – One chip multiprocessor
• AMD and INTEL- Two –Processor
• SUN – 8 processor multi core
Symmetric shared – memory support caching of
• Shared Data
• Private Data
Private data: used by a single processor
When a private item is cached, its location is migrated to the cache Since no other
processor uses the data, the program behavior is identical to that in a uniprocessor.

Shared data: used by multiple processor

When shared data are cached, the shared value may be replicated in multiple caches
advantages: reduce access latency and memory contention induces a new problem: cache
coherence.

Cache Coherence
Unfortunately, caching shared data introduces a new problem because the view of
memory held by two different processors is through their individual caches, which,
without any additional precautions, could end up seeing two different values.
I.e, If two different processors have two different values for the same location, this
difficulty is generally referred to as cache coherence problem
 Cache coherence problem for a single memory location
• Informally:
– “Any read must return the most recent write”
– Too strict and too difficult to implement
• Better:
– “Any write must eventually be seen by a read”
– All writes are seen in proper order (“serialization”)
• Two rules to ensure this:
– “If P writes x and then P1 reads it, P’s write will be seen by P1 if the read
and write are sufficiently far apart”
– Writes to a single location are serialized: seen in one order
• Latest write will be seen
• Otherwise could see writes in illogical order (could see older
value after a newer value)

The definition contains two different aspects of memory system:

• Coherence
• Consistency
A memory system is coherent if,
• Program order is preserved.
• Processor should not continuously read the old data value.
• Write to the same location are serialized.
The above three properties are sufficient to ensure coherence,When a written value will
be seen is also important. This issue is defined by memory consistency model. Coherence
and consistency are complementary.
Basic schemes for enforcing coherence
Coherence cache provides:
• migration: a data item can be moved to a local cache and used there in a
transparent fashion.
• replication for shared data that are being simultaneously read.
• both are critical to performance in accessing shared data.
To over come these problems, adopt a hardware solution by introducing a protocol to
maintain coherent caches named as Cache Coherence Protocols
These protocols are implemented for tracking the state of any sharing of a data block.
Two classes of Protocols
• Directory based
• Snooping based

Directory based
• Sharing status of a block of physical memory is kept in one location called the
directory.
• Directory-based coherence has slightly higher implementation overhead than
snooping.
• It can scale to larger processor count.

Snooping
• Every cache that has a copy of data also has a copy of the sharing status of the
block.
• No centralized state is kept.
• Caches are also accessible via some broadcast medium (bus or switch)
• Cache controller monitor or snoop on the medium to determine whether or not
they have a copy of a block that is represented on a bus or switch access.
Snooping protocols are popular with multiprocessor and caches attached to single shared
memory as they can use the existing physical connection- bus to memory, to interrogate
the status of the caches. Snoop based cache coherence scheme is implemented on a
shared bus. Any communication medium that broadcasts cache misses to all the
processors.

Basic Snoopy Protocols

• Write strategies
– Write-through: memory is always up-to-date
– Write-back: snoop in caches to find most recent copy

• Write Invalidate Protocol

– Multiple readers, single writer
– Write to shared data: an invalidate is sent to all caches which snoop and
invalidate any copies
• Read miss: further read will miss in the cache and fetch a new
copy of the data.

• Write Broadcast/Update Protocol (typically write through)

– Write to shared data: broadcast on bus, processors snoop, and update any
copies
– Read miss: memory/cache is always up-to-date.
• Write serialization: bus serializes requests!
– Bus is single point of arbitration

Examples of Basic Snooping Protocols

Write Invalidate
Write Update

Assume neither cache initially holds X and the value of X in memory is 0

Example Protocol
• Snooping coherence protocol is usually implemented by incorporating a finite-
state controller in each node
• Logically, think of a separate controller associated with each cache block
– That is, snooping operations or cache requests for different blocks can
proceed independently
• In implementations, a single controller allows multiple operations to distinct
blocks to proceed in interleaved fashion
 – that is, one operation may be initiated before another is completed, even
through only one cache access or one bus access is allowed at time
Example Write Back Snoopy Protocol

• Invalidation protocol, write-back cache

– Snoops every address on bus
– If it has a dirty copy of requested block, provides that block in response to
the read request and aborts the memory access
• Each memory block is in one state:
– Clean in all caches and up-to-date in memory (Shared)
– OR Dirty in exactly one cache (Exclusive)
– OR Not in any caches
• Each cache block is in one state (track these):
– Shared : block can be read
– OR Exclusive : cache has only copy, its writeable, and dirty
– OR Invalid : block contains no data (in uniprocessor cache too)
• Read misses: cause all caches to snoop bus
• Writes to clean blocks are treated as misses

Write-Back State Machine – CPU

State Transitions for Each Cache Block is as shown below
• CPU may read/write hit/miss to the block
• May place write/read miss on bus
• May receive read/write miss from bus
Cache Coherence State Diagram
Conclusion
• “End” of uniprocessors speedup => Multiprocessors
• Parallelism challenges: % parallalizable, long latency to remote memory
• Centralized vs. distributed memory
– Small MP vs. lower latency, larger BW for Larger MP
• Message Passing vs. Shared Address
– Uniform access time vs. Non-uniform access time
• Snooping cache over shared medium for smaller MP by invalidating other cached
copies on write
• Sharing cached data ⇒ Coherence (values returned by a read), Consistency
(when a written value will be returned by a read)
• Shared medium serializes writes ⇒ Write consistency
Implementation Complications
• Write Races:
– Cannot update cache until bus is obtained
• Otherwise, another processor may get bus first,
and then write the same cache block!
– Two step process:
• Arbitrate for bus
• Place miss on bus and complete operation
– If miss occurs to block while waiting for bus,
handle miss (invalidate may be needed) and then restart.
– Split transaction bus:
• Bus transaction is not atomic:
can have multiple outstanding transactions for a block
• Multiple misses can interleave,
allowing two caches to grab block in the Exclusive state
• Must track and prevent multiple misses for one block
• Must support interventions and invalidations

Performance Measurement
• Overall cache performance is a combination of
– Uniprocessor cache miss traffic
– Traffic caused by communication – invalidation and subsequent cache
misses
• Changing the processor count, cache size, and block size can affect these two
components of miss rate
• Uniprocessor miss rate: compulsory, capacity, conflict
• Communication miss rate: coherence misses
– True sharing misses + false sharing misses
True and False Sharing Miss
• True sharing miss
– The first write by a PE to a shared cache block causes an invalidation to
establish ownership of that block
– When another PE attempts to read a modified word in that cache block, a
miss occurs and the resultant block is transferred
• False sharing miss
– Occur when a block a block is invalidate (and a subsequent reference
causes a miss) because some word in the block, other than the one being
read, is written to
– The block is shared, but no word in the cache is actually shared, and
this miss would not occur if the block size were a single word
• Assume that words x1 and x2 are in the same cache block, which is in the shared
state in the caches of P1 and P2. Assuming the following sequence of events,
identify each miss as a true sharing miss or a false sharing miss.

Time P1 P2
1 Write x1
2 Read x2
3 Write x1
4 Write x2
5 Read x2

Example Result
• True sharing miss (invalidate P2)
• 2: False sharing miss
– x2 was invalidated by the write of P1, but that value of x1 is not used in
P2
• 3: False sharing miss
 – The block containing x1 is marked shared due to the read in P2, but P2 did
not read x1. A write miss is required to obtain exclusive access to the
block
• 4: False sharing miss
• 5: True sharing miss

Distributed Shared-Memory Architectures

Distributed shared-memory architectures
• Separate memory per processor
– Local or remote access via memory controller
– The physical address space is statically distributed
Coherence Problems
• Simple approach: uncacheable
– shared data are marked as uncacheable and only private data are kept in
caches
– very long latency to access memory for shared data
• Alternative: directory for memory blocks
– The directory per memory tracks state of every block in every cache
• which caches have a copies of the memory block, dirty vs. clean, ...
– Two additional complications
• The interconnect cannot be used as a single point of arbitration like
the bus
• Because the interconnect is message oriented, many messages
must have explicit responses
To prevent directory becoming the bottleneck, we distribute directory entries with
memory, each keeping track of which processors have copies of their memory blocks
Directory Protocols
• Similar to Snoopy Protocol: Three states
– Shared: 1 or more processors have the block cached, and the value in
memory is up-to-date (as well as in all the caches)
 – Uncached: no processor has a copy of the cache block (not valid in any
cache)
– Exclusive: Exactly one processor has a copy of the cache block, and it has
written the block, so the memory copy is out of date
• The processor is called the owner of the block

• In addition to tracking the state of each cache block, we must track the processors
that have copies of the block when it is shared (usually a bit vector for each
memory block: 1 if processor has copy)

• Keep it simple(r):
– Writes to non-exclusive data => write miss
– Processor blocks until access completes
– Assume messages received and acted upon in order sent

Messages for Directory Protocols

• local node: the node where a request originates

• home node: the node where the memory location and directory entry of an address
reside
• remote node: the node that has a copy of a cache block (exclusive or shared)
State Transition Diagram for Individual Cache Block

• Comparing to snooping protocols:

– identical states
– stimulus is almost identical
– write a shared cache block is treated as a write miss (without fetch the
block)
– cache block must be in exclusive state when it is written
– any shared block must be up to date in memory
• write miss: data fetch and selective invalidate operations sent by the directory
controller (broadcast in snooping protocols)
Directory Operations: Requests and Actions
• Message sent to directory causes two actions:
– Update the directory
– More messages to satisfy request
• Block is in Uncached state: the copy in memory is the current value; only possible
requests for that block are:
– Read miss: requesting processor sent data from memory &requestor made
only sharing node; state of block made Shared.
– Write miss: requesting processor is sent the value & becomes the Sharing
node. The block is made Exclusive to indicate that the only valid copy is
cached. Sharers indicates the identity of the owner.
• Block is Shared => the memory value is up-to-date:
– Read miss: requesting processor is sent back the data from memory &
requesting processor is added to the sharing set.
– Write miss: requesting processor is sent the value. All processors in the set
Sharers are sent invalidate messages, & Sharers is set to identity of
requesting processor. The state of the block is made Exclusive.
• Block is Exclusive: current value of the block is held in the cache of the processor
identified by the set Sharers (the owner) => three possible directory requests:
– Read miss: owner processor sent data fetch message, causing state of
block in owner’s cache to transition to Shared and causes owner to send
data to directory, where it is written to memory & sent back to requesting
processor.
Identity of requesting processor is added to set Sharers, which still
contains the identity of the processor that was the owner (since it still has a
readable copy). State is shared.
– Data write-back: owner processor is replacing the block and hence must
write it back, making memory copy up-to-date
(the home directory essentially becomes the owner), the block is now
Uncached, and the Sharer set is empty.
– Write miss: block has a new owner. A message is sent to old owner
causing the cache to send the value of the block to the directory from
which it is sent to the requesting processor, which becomes the new owner.
Sharers is set to identity of new owner, and state of block is made
Exclusive.
Synchronization : The Basics
Synchronization mechanisms are typically built with user-level software routines that rely
on hardware –supplied synchronization instructions.
• Why Synchronize?
Need to know when it is safe for different processes to use shared data
• Issues for Synchronization:
– Uninterruptable instruction to fetch and update memory (atomic
operation);
– User level synchronization operation using this primitive;
– For large scale MPs, synchronization can be a bottleneck; techniques to
reduce contention and latency of synchronization

Uninterruptable Instruction to Fetch and Update Memory

• Atomic exchange: interchange a value in a register for a value in memory
0 ⇒ synchronization variable is free
1 ⇒ synchronization variable is locked and unavailable
– Set register to 1 & swap
– New value in register determines success in getting lock
0 if you succeeded in setting the lock (you were first)
1 if other processor had already claimed access
– Key is that exchange operation is indivisible
• Test-and-set: tests a value and sets it if the value passes the test
• Fetch-and-increment: it returns the value of a memory location and atomically
increments it
– 0 ⇒ synchronization variable is free
• Hard to have read & write in 1 instruction: use 2 instead
• Load linked (or load locked) + store conditional
– Load linked returns the initial value
– Store conditional returns 1 if it succeeds (no other store to same memory
location since preceding load) and 0 otherwise
• Example doing atomic swap with LL & SC:
 try: mov R3,R4 ; mov exchange value
ll R2,0(R1) ; load linked
sc R3,0(R1) ; store conditional
beqz R3,try ; branch store fails (R3 = 0)
mov R4,R2 ; put load value in R4
• Example doing fetch & increment with LL & SC:
try: ll R2,0(R1) ; load linked
addi R2,R2,#1 ; increment (OK if reg–reg)
sc R2,0(R1) ; store conditional
beqz R2,try ; branch store fails (R2 = 0)

User Level Synchronization—Operation Using this Primitive

• Spin locks: processor continuously tries to acquire, spinning around a loop trying
to get the lock
li R2,#1
lockit: exch R2,0(R1) ;atomic exchange
bnez R2,lockit ;already locked?
• What about MP with cache coherency?
– Want to spin on cache copy to avoid full memory latency
– Likely to get cache hits for such variables
• Problem: exchange includes a write, which invalidates all other copies; this
generates considerable bus traffic
• Solution: start by simply repeatedly reading the variable; when it changes, then
try exchange (“test and test&set”):
try: li R2,#1
lockit: lw R3,0(R1) ;load var
bnez R3,lockit ;≠ 0 ⇒ not free ⇒ spin
exch R2,0(R1) ;atomic exchange
bnez R2,try ;already locked?
Memory Consistency Models
• What is consistency? When must a processor see the new value? e.g., seems that
P1: A = 0; P2: B = 0;
..... .....
A = 1; B = 1;
L1: if (B == 0) ... L2: if (A == 0) ...
• Impossible for both if statements L1 & L2 to be true?
– What if write invalidate is delayed & processor continues?
• Memory consistency models:
what are the rules for such cases?
• Sequential consistency: result of any execution is the same as if the accesses of
each processor were kept in order and the accesses among different processors
were interleaved ⇒ assignments before ifs above
– SC: delay all memory accesses until all invalidates done
• Schemes faster execution to sequential consistency
• Not an issue for most programs; they are synchronized
– A program is synchronized if all access to shared data are ordered by
synchronization operations
write (x)
...
release (s) {unlock}
...
acquire (s) {lock}
...
read(x)
• Only those programs willing to be nondeterministic are not synchronized: “data
race”: outcome f(proc. speed)
• Several Relaxed Models for Memory Consistency since most programs are
synchronized; characterized by their attitude towards: RAR, WAR, RAW, WAW
to different addresses
Relaxed Consistency Models : The Basics
• Key idea: allow reads and writes to complete out of order, but to use
synchronization operations to enforce ordering, so that a synchronized program
behaves as if the processor were sequentially consistent
– By relaxing orderings, may obtain performance advantages
– Also specifies range of legal compiler optimizations on shared data
– Unless synchronization points are clearly defined and programs are
synchronized, compiler could not interchange read and write of 2 shared
data items because might affect the semantics of the program
• 3 major sets of relaxed orderings:
1. W→R ordering (all writes completed before next read)
• Because retains ordering among writes, many programs that operate under
sequential consistency operate under this model, without additional
synchronization. Called processor consistency
2. W → W ordering (all writes completed before next write)
3. R → W and R → R orderings, a variety of models depending on ordering
restrictions and how synchronization operations enforce ordering
• Many complexities in relaxed consistency models; defining precisely what it
means for a write to complete; deciding when processors can see values that it has
written
 Chapter-7
Memory Hierarchy Design
• Unlimited amount of fast memory
- Economical solution is memory hierarchy
- Locality
- Cost performance
Principle of locality
- most programs do not access all code or data uniformly.
• Locality occurs
- Time (Temporal locality)
- Space (spatial locality)
• Guidelines
– Smaller hardware can be made faster
– Different speed and sizes

Goal is provide a memory system with cost per byte than the next lower level
• Each level maps addresses from a slower, larger memory to a smaller but faster
memory higher in the hierarchy.
– Address mapping
 – Address checking.
• Hence protection scheme for address for scrutinizing addresses are
also part of the memory hierarchy.
Memory Hierarchy

Levels of the Memory Hierarchy Upper

Capacity Faste
Access Time

CPU Registers
500 bytes Register
0.25 ns
Cache
64 KB
Cach

Capacit
Spee
1 ns

Block
Main Memory
512 MB Memor
100ns
Page
Disk
100 GB I/O
5 ms
File Large
??? Lower

Why More on Memory Hierarchy?

100,000

10,000
Performance

1,000
Processor
100

10
Memory

1
1980 1985 1990 1995 2000 2005 2010
Year
 • The importance of memory hierarchy has increased with advances in performance
of processors.
• Prototype
– When a word is not found in cache
• Fetched from memory and placed in cache with the address tag.
• Multiple words( block) is fetched for moved for efficiency reasons.

– key design
• Set associative
– Set is a group of block in the cache.
– Block is first mapped on to set.
» Find mapping
» Searching the set
Chosen by the address of the data:
(Block address) MOD(Number of sets in cache)
• n-block in a set
– Cache replacement is called n-way set associative.
Cache data
- Cache read.
- Cache write.
Write through: update cache and writes through to update memory.
Both strategies
- Use write buffer.
this allows the cache to proceed as soon as the data is placed in the
buffer rather than wait the full latency to write the data into memory.
Metric
used to measure the benefits is miss rate
No of access that miss
No of accesses
Write back: updates the copy in the cache.
• Causes of high miss rates
 – Three Cs model sorts all misses into three categories
• Compulsory: every first access cannot be in cache
– Compulsory misses are those that occur if there is an
infinite cache
• Capacity: cache cannot contain all that blocks that are needed for
the program.
– As blocks are being discarded and later retrieved.
• Conflict: block placement strategy is not fully associative
– Block miss if blocks map to its set.

Miss rate can be a misleading measure for several reasons

So, misses per instruction can be used per memory reference
Misses = Miss rate X Memory accesses
Instruction Instruction count

= Miss rate X Mem accesses

Instruction
Cache Optimizations
Six basic cache optimizations
1. Larger block size to reduce miss rate:
- To reduce miss rate through spatial locality.
- Increase block size.
- Larger block size reduce compulsory misses.
- But they increase the miss penalty.
2. Bigger caches to reduce miss rate:

- capacity misses can be reduced by increasing the cache capacity.

- Increases larger hit time for larger cache memory and higher cost and power.

3. Higher associativity to reduce miss rate:

- Increase in associativity reduces conflict misses.
4. Multilevel caches to reduce penalty:
- Introduces additional level cache
- Between original cache and memory.
- L1- original cache
L2- added cache.
L1 cache: - small enough
- speed matches with clock cycle time.
L2 cache: - large enough
- capture many access that would go to main memory.
Average access time can be redefined as
Hit timeL1+ Miss rate L1 X ( Hit time L2 + Miss rate L2 X Miss penalty L2)
5. Giving priority to read misses over writes to reduce miss penalty:
- write buffer is a good place to implement this optimization.
- write buffer creates hazards: read after write hazard.
6. Avoiding address translation during indexing of the cache to reduce hit time:
- Caches must cope with the translation of a virtual address from the processor to a
physical address to access memory.
- common optimization is to use the page offset.
- part that is identical in both virtual and physical addresses- to index the cache.

Advanced Cache Optimizations

• Reducing hit time
– Small and simple caches
– Way prediction
– Trace caches
• Increasing cache bandwidth
– Pipelined caches
– Multibanked caches
– Nonblocking caches
• Reducing Miss Penalty
– Critical word first
– Merging write buffers
• Reducing Miss Rate
– Compiler optimizations
• Reducing miss penalty or miss rate via parallelism
– Hardware prefetching
– Compiler prefetching
First Optimization : Small and Simple Caches
• Index tag memory and then compare takes time
• ⇒ Small cache can help hit time since smaller memory takes less time to index
– E.g., L1 caches same size for 3 generations of AMD microprocessors: K6,
Athlon, and Opteron
– Also L2 cache small enough to fit on chip with the processor avoids time
penalty of going off chip
• Simple ⇒ direct mapping
– Can overlap tag check with data transmission since no choice
• Access time estimate for 90 nm using CACTI model 4.0
– Median ratios of access time relative to the direct-mapped caches are 1.32,
1.39, and 1.43 for 2-way, 4-way, and 8-way caches

2.50
Access time (ns)

2.00 1-way 2-way 4-way 8-way

1.50

1.00
0.50
-
16 KB 32 KB 64 KB 128 KB 256 KB 512 KB 1 MB
Cache size

Second Optimization: Way Prediction

• How to combine fast hit time of Direct Mapped and have the lower conflict
misses of 2-way SA cache?
• Way prediction: keep extra bits in cache to predict the “way,” or block within the
set, of next cache access.

Hit Time

Way-Miss Hit Time Miss Penalty

– Multiplexer is set early to select desired block, only 1 tag comparison

performed that clock cycle in parallel with reading the cache data
 – Miss ⇒ 1st check other blocks for matches in next clock cycle
• Accuracy ≈ 85%
• Drawback: CPU pipeline is hard if hit takes 1 or 2 cycles
– Used for instruction caches vs. data caches
Third optimization: Trace Cache
• Find more instruction level parallelism?
How to avoid translation from x86 to microops?
• Trace cache in Pentium 4
1. Dynamic traces of the executed instructions vs. static sequences of instructions as
determined by layout in memory
– Built-in branch predictor
2. Cache the micro-ops vs. x86 instructions
– Decode/translate from x86 to micro-ops on trace cache miss
+ 1. ⇒ better utilize long blocks (don’t exit in middle of block, don’t enter at label
in middle of block)
- 1. ⇒ complicated address mapping since addresses no longer aligned to power-
of-2 multiples of word size
- 1. ⇒ instructions may appear multiple times in multiple dynamic traces due to
different branch outcomes
Fourth optimization: pipelined cache access to increase bandwidth
• Pipeline cache access to maintain bandwidth, but higher latency
• Instruction cache access pipeline stages:
1: Pentium
2: Pentium Pro through Pentium III
4: Pentium 4
- ⇒ greater penalty on mispredicted branches
- ⇒ more clock cycles between the issue of the load and the use of the data
Fifth optimization: Increasing Cache Bandwidth Non-Blocking Caches
• Non-blocking cache or lockup-free cache allow data cache to continue to supply
cache hits during a miss
– requires F/E bits on registers or out-of-order execution
– requires multi-bank memories
• “hit under miss” reduces the effective miss penalty by working during miss vs.
ignoring CPU requests
• “hit under multiple miss” or “miss under miss” may further lower the effective
miss penalty by overlapping multiple misses
– Significantly increases the complexity of the cache controller as there can
be multiple outstanding memory accesses
– Requires multiple memory banks (otherwise cannot support)
– Pentium Pro allows 4 outstanding memory misses
Value of Hit Under Miss for SPEC

Hit Under i Misses

1.8

1.6

1.4
0->1
1.2
1->2
1
2->64
0.8
Base
0.6

0.4

0.2

0
doduc

nasa7
espresso

ear

ora
wave5
eqntott

compress

tomcatv
fpppp

su2cor

hydro2d

spice2g6
xlisp

alvinn
swm256

mdljdp2
mdljsp2

Integer Floating Point

• FP programs on average: AMAT= 0.68 -> 0.52 -> 0.34 -> 0.26
 • Int programs on average: AMAT= 0.24 -> 0.20 -> 0.19 -> 0.19
• 8 KB Data Cache, Direct Mapped, 32B block, 16 cycle miss, SPEC 92

Sixth optimization: Increasing Cache Bandwidth via Multiple Banks

• Rather than treat the cache as a single monolithic block, divide into independent
banks that can support simultaneous accesses
– E.g.,T1 (“Niagara”) L2 has 4 banks
• Banking works best when accesses naturally spread themselves across banks ⇒
mapping of addresses to banks affects behavior of memory system
• Simple mapping that works well is “sequential interleaving”
– Spread block addresses sequentially across banks
– E,g, if there 4 banks, Bank 0 has all blocks whose address modulo 4 is 0;
bank 1 has all blocks whose address modulo 4 is 1; …
Seventh optimization :Reduce Miss Penalty: Early Restart and Critical
Word First
• Don’t wait for full block before restarting CPU
• Early restart—As soon as the requested word of the block arrives, send it to the
CPU and let the CPU continue execution
– Spatial locality ⇒ tend to want next sequential word, so not clear size of
benefit of just early restart
• Critical Word First—Request the missed word first from memory and send it to
the CPU as soon as it arrives; let the CPU continue execution while filling the rest
of the words in the block
– Long blocks more popular today ⇒ Critical Word 1st Widely used

block

Eight optimization: Merging Write Buffer to Reduce Miss Penalty-

• Write buffer to allow processor to continue while waiting to write to memory
• If buffer contains modified blocks, the addresses can be checked to see if address
of new data matches the address of a valid write buffer entry
• If so, new data are combined with that entry
• Increases block size of write for write-through cache of writes to sequential words,
bytes since multiword writes more efficient to memory
• The Sun T1 (Niagara) processor, among many others, uses write merging
Ninth optimization: Reducing Misses by Compiler Optimizations
• McFarling [1989] reduced caches misses by 75% on 8KB direct mapped cache, 4
byte blocks in software
• Instructions
– Reorder procedures in memory so as to reduce conflict misses
– Profiling to look at conflicts (using tools they developed)
• Data
– Merging Arrays: improve spatial locality by single array of compound
elements vs. 2 arrays
– Loop Interchange: change nesting of loops to access data in order stored in
memory
– Loop Fusion: Combine 2 independent loops that have same looping and
some variables overlap
 – Blocking: Improve temporal locality by accessing “blocks” of data
repeatedly vs. going down whole columns or rows
Merging Arrays Example
/* Before: 2 sequential arrays */
int val[SIZE];
int key[SIZE];

/* After: 1 array of stuctures */

struct merge {
int val;
int key;
};
struct merge merged_array[SIZE];
Reducing conflicts between val & key; improve spatial locality

0.1

Direct Mapped Cache

0.05

Fully Associative Cache

0
0 50 100 150
Blocking Factor

• Conflict misses in caches not FA vs. Blocking size

– Lam et al [1991] a blocking factor of 24 had a fifth the misses vs. 48
despite both fit in cache
Tenth optimization Reducing Misses by Hardware Prefetching of
Instructions & Data
• Prefetching relies on having extra memory bandwidth that can be used without
penalty
• Instruction Prefetching
– Typically, CPU fetches 2 blocks on a miss: the requested block and the
next consecutive block.
– Requested block is placed in instruction cache when it returns, and
prefetched block is placed into instruction stream buffer
• Data Prefetching
– Pentium 4 can prefetch data into L2 cache from up to 8 streams from 8
different 4 KB pages
– Prefetching invoked if 2 successive L2 cache misses to a page,
if distance between those cache blocks is < 256 bytes
Performance Improvement

2.20 1.97
2.00
1.80
1.60 1.45 1.49
1.40
1.26 1.29 1.32
1.40 1.18 1.20 1.21
1.16
1.20
1.00
u
p

el
3d

ke
im
c

id
cf

s
re

pl

ca
is
ga

lg

gr
m

ua
sw
m

pw

ap
ce
ga

lu

m
fa

eq
fa
wu

SPECint2000 SPECfp2000

Eleventh optimization: Reducing Misses by Software Prefetching Data

• Data Prefetch
– Load data into register (HP PA-RISC loads)
– Cache Prefetch: load into cache
(MIPS IV, PowerPC, SPARC v. 9)
– Special prefetching instructions cannot cause faults;
a form of speculative execution
• Issuing Prefetch Instructions takes time
 – Is cost of prefetch issues < savings in reduced misses?
– Higher superscalar reduces difficulty of issue bandwidth

The techniques to improve hit time, bandwidth, miss penalty and miss rate generally
affect the other components of the average memory access equation as well as the
complexity of the memory hierarchy.