# Cache hierarchy

Generic Multi-Level Cache Organization

Cache hierarchy, or multi-level caches, refers to a memory model designed generally to hold data which is more likely to be requested by processors. The purpose of these memory models is to provide a faster execution of memory related instructions, and faster overall performance of the system.

This model was for CPU cores to run at faster clocks that needed to hide the memory latency of the main memory access. Today Multi-level caches are the optimum solution to provide fast access to data residing in the main memory. The access time to memory that acts as a bottleneck for the CPU core performance can be relaxed by using a hierarchical cache structure in order to reduce the latency and therefore speed up the CPU clock.[1]

## Background

In the history of computer and electronic chip developments, there was a time that CPUs were getting faster while memory access speeds had not yet caught up. At the time, the gap and difference between CPUs and memories represented a need for enhancements in memory access time. With CPUs becoming faster, systems were more capable of running and executing more instructions in a given time, but the time limitation in data access from memory prevented programmers from benefitting from this capability. This issue was the motivation behind thoughts for achieving memory models with a higher access rate in order to work with processors to achieve better and faster performances. Therefore, the need for such memory models resulted in the concept of Cache memory. This concept was first proposed by Maurice Wilkes, a British computer scientist at the University of Cambridge in 1965, but at the time he called such memories - "slave memory".[2] Roughly between 1970-1990 there were lots of papers and articles proposed by many people like Anant Agarwal, Alan Jay Smith, Mark D. Hill, Thomas R. Puzak, etc., regarding enhancement and analysis for better cache memory designs. The first cache memory models were implemented at the time, but as researchers were investigating and proposing better designs, the need for faster memory models was still present; the reason being that although those cache models improved data access latency, they could not have enough storage capacity to cover as much data compared to the size of the main memory and there was a lot of data to be accessed in the old fashioned way with high latency. Therefore, approximately from 1990 and so on, gradually ideas like adding another cache level (second-level) to such memory models as a backup for the first level cache came into thoughts and proposals. Many people, including Jean-Loup Baer, Wen-Hann Wang, Andrew W. Wilson, etc. have conducted researches on this model. When several simulations and implementations demonstrated the advantages of two-level cache models having a faster data access from memory, the concept of multi-level caches formed as a new and generally better model of cache memories compared to its single form. From the year 2000 until now, multi-level cache models received a widespread attention and are implemented wildly in many systems, as three-level caches are present in Intel Core i7 products.[3]

## Multi-level cache

Access to the main memory for each instruction execution may result in slow processing, with the clock speed dependent on the time taken to find the data in the main memory and fetch it. In order to hide this memory latency from the processor, data caching is used. Whenever the data is required by the processor, it is fetched from the memory and stored in the small structure called Cache. For any further references to that data, the cache is searched first before going to the main memory. This structure resides closer to the processor in terms of time taken to search and fetch data with respect to the main memory. The advantages of using cache can be proven by calculating the Average Access Time (AAT) for the memory hierarchy without cache and with the cache.

### Average access time (AAT)

Caches, being small in size, may result in frequent misses and the computer might eventually have to go to the main memory to fetch data. And hence, the AAT depends on the miss rate of all the structures that it searches through for the data.[4]

${\displaystyle {\text{AAT}}={\text{hit time}}+(({\text{miss rate}})\times ({\text{miss penalty}}))}$

AAT for main memory is given by Hit time main memory. AAT for caches can be given by

Hit timecache + (Miss ratecache + Miss Penaltytime taken to go to main memory after missing cache).

Hit time for caches is much less than the hit time for the main memory and hence the resulting AAT after using cache in the memory hierarchy is improved significantly.

While using the cache to improve memory latency, it may not always result in the required improvement for the time taken to fetch data due to the way caches are organized and traversed. For example, direct-mapped caches that are the same size usually have a higher miss rate than the fully associative caches. This may also depend on upon the benchmark that the computer testing the processor upon and the pattern of instructions. But always using the fully associative cache may result in more power consumption as it has to search the whole cache every time. Due to this, the trade-off between the power consumption and the size of the cache becomes critical in the cache design.[4]

### Evolution

In the case of a miss in the cache, the purpose of using such a structure will be rendered useless and the computer will ultimately have to go to the main memory to fetch the required data. The idea of using multiple levels of cache comes into picture here. This means that if the computer misses the cache closest to the processor, the computer will search for the data in the next closest level of cache and will continue until the computer run out of levels of caches and will finally search the main memory. The general trend is to keep the L1 cache small and at a distance of 1–2 CPU clock cycles from the processor with the lower levels of caches increasing in size to store more data than L1 and hence have a lower miss rate. This, in turn, results into a better AAT. The number of levels of cache can be designed by architects as per the requirements after checking for the trade-offs between cost, AATs, and size.

Cache hierarchy for up-to L3 level of cache and main memory with on chip L1.

### Performance gains

With the technology scaling which made memory systems feasible and smaller to accommodate on a single chip, most of the modern day processors go for up to 3 or 4 levels of caches. The reduction in the AAT can be understood by this example, where the computer checks AAT for different configurations up to 3-level caches.

Example: Main memory = 50ns, L1 = 1ns (10% miss rate), L2 = 5ns (1% miss rate), L3 = 10 ns (0.2% miss rate) ·

1. AAT (No cache) = 50ns
2. AAT (L1 cache+ Main Memory) = 1ns + (0.1 × 50ns) = 6ns
3. AAT (L1 cache + L2 cache + Main Memory) = 1ns + (0.1 × (5 + 0.01(50ns)) = 1.55ns
4. AAT (L1 cache + L2 cache + L3 cache + Main Memory) = 1ns + (0.1 × (5 + 0.01(10 + 0.002 × 50ns))) = 1.5001ns

• Increase the cost of memories and hence the overall system.
• Cached data is stored only so long as power supply is provided.
• Increase in area consumed by memory system on chip.[5]
• In case of a large programs with poor temporal locality, even the multi-level caches cannot help improve the performance and eventually, the main memory needs to be reached to fetch the data.[6]

## Properties

Cache organization with L1 cache as Separate and L2 cache as Unified

### Banked versus unified

In a banked cache, the cache is divided into an instruction cache and a data cache. In contrast, a unified cache contains both the instructions and data combined in the same cache. During a process, the upper-level cache is accessed to get the instructions to the processor in each cycle. The cache will also be accessed to get data. Requiring both actions to be implemented at the same time requires multiple ports in the cache and more access time. Having multiple ports requires additional hardware and wiring, leading to huge structure. Therefore, the L1 cache is organized as a banked cache which results in fewer ports, less hardware, and low access time.[4]

The lower level caches L2 and L3 are accessed only when there is a miss in the L1 cache implies the frequency of access to the lower level caches is less compared to the L1 cache. Therefore, the unified organization is implemented in the lower level caches, as having a single port will suffice.

### Inclusion policies

Inclusive cache organization

Whether a block present in the upper cache layer can be present in the lower cache level is governed by the inclusion policies below:[7]

• Inclusive
• Exclusive
• Non-Inclusive Non-Exclusive (NINE)

In the Inclusive policy, all the blocks present in the upper-level cache have to be present in the lower-level cache as well. Each upper-level cache component is a subset of the lower-level cache component. In this case, since there is a duplication of blocks, there is some wastage of memory. However, checking is better in the case of inclusive, because if the lower-level cache doesn't have the block, then the computer can be sure that the upper-level cache can no way have that block.[7]

In the exclusive policy, all the cache hierarchy components are completely exclusive, which implies that any element in the upper-level cache will not be present in any of the lower cache component. This enables complete usage of the cache memory as no same block is present in the other cache components. However, there is a high memory access latency.[8]

The above policies require a set on rules to be followed in order to implement them. If none of these are forced, the resulting inclusion policy is called Non-Inclusive Non-Exclusive (NINE). This means that the upper-level cache may or may not be present in the lower level cache.[6]

### Write policies

There are two policies which define the way in which a modified cache block will be updated in the main memory:[7]

• Write Through
• Write Back

In the case of Write Through policy whenever the value of the cache block changes, it is further modified in the lower-level memory hierarchy as well. This policy ensures that the data is stored safely as it is written throughout the hierarchy.

However, in the case of the write back policy, the changed cache block will be updated in the lower-level hierarchy only when the cache block is evicted. Writing back every cache block which is evicted is not efficient. Therefore, the computer use the concept of a Dirty bit attached to each cache block. The dirty bit is made high whenever the cache block is modified and during eviction, only the blocks with dirty bit high will be written to the lower-level hierarchy and then the dirty bit is cleared. In this policy, there is data losing risk as the only valid copy is stored in the cache and therefore need some correction techniques to be implemented.

In case of a write where the byte is not present in the cache block the write policies below determine whether the byte has to be brought to the cache or not:[7]

• Write Allocate
• Write No-Allocate

Write Allocate policy states that in case of a write miss, the block is fetched from the main memory and placed in the cache before writing. In the Write No-Allocate policy, if the block is missed in the cache it will just write in the lower level memory hierarchy without fetching the block into the cache.

The common combinations of the policies are Write Back Write Allocate and Write Through Write No-Allocate.

Cache organization with L1 as private and L2 and L3 as shared

### Shared versus private

A private cache is private to that particular core and cannot be accessed by the other cores. Since each core has its own private cache there might be duplicate blocks in the cache which leads to reduced capacity utilization. However, this organization leads to a lower latency.[7][9][10]

A shared cache is where it is shared among multiple cores and therefore can be directly accessed by any of the cores. Since it is shared, each block in the cache is unique and therefore has more hit rate as there will be no duplicate blocks. However, the cache hit latency is larger as multiple cores try to access the same cache.

In the multi-core processors, the organization of the cache to be shared or private impacts the performance of the processor. In practice, the upper-level cache L1 (or sometimes L2)[11][12] is implemented as private and lower level caches are implemented as shared.

## Recent implementation models

[13] Cache Organization of Intel Nehalem Microarchitecture

### Intel Broadwell Microarchitecture (2014)

• L1 Cache (Instruction and Data) – 64kB per core
• L2 Cache – 256kB per core
• L3 Cache – 2MB to 6MB shared
• L4 Cache – 128MB of eDRAM (Iris Pro Models only)[11]

### Intel Kaby Lake Microarchitecture (2016)

• L1 Cache (Instruction and Data) – 64kB per core
• L2 Cache – 256kB per core
• L3 Cache – 2MB to 8MB shared[12]

### IBM Power 7

• L1 Cache (Instruction and Data) – Each 64-banked, each bank has 2rd+1wr ports 32kB, 8-way associative, 128B block, Write through
• L2 Cache – 256kB, 8-way, 128B block, Write back, Inclusive of L1, 2ns access latency
• L3 Cache – 8 regions of 4MB (total 32MB), local region 6ns, remote 30ns, each region 8-way associative, DRAM data array, SRAM tag array[14]

## References

1. ^ "Cache: Why Level It" (PDF).
2. ^ "Sir Maurice Vincent Wilkes | British computer scientist". Encyclopædia Britannica. Retrieved 2016-12-11.
3. ^ Berkeley, John L. Hennessy, Stanford University, and David A. Patterson, University of California,. "Memory Hierarchy Design - Part 6. The Intel Core i7, fallacies, and pitfalls". EDN. Retrieved 2016-12-11.
4. ^ a b c Hennessey and Patterson. Computer Architecture: A Quantitative Approach. Morgan Kaufmann. ISBN 9780123704900.
5. ^
6. ^ a b Solihin, Yan (2016). Fundamentals of Parallel Multicore Architecture. Chapman and Hall. pp. Chapter 5: Introduction to Memory Hierarchy Organization. ISBN 9781482211184.
7. Solihin, Yan (2009). Fundamentals of Parallel Computer Architecture. Solihin Publishing. pp. Chapter 6: Introduction to Memory Hierarchy Organization. ISBN 9780984163007.
8. ^
9. ^
10. ^
11. ^ a b
12. ^ a b
13. ^
14. ^