By Hongbin Sun, Nanning Zheng , and Tong Zhang

LEVERAGING ACCESS LOCALITY FOR THE EFFICIENT USE

OF MULTIBIT ERROR-CORRECTING CODES IN L2 CACHE

By Hongbin Sun, Nanning Zheng, and Tong Zhang

Joseph SchneiderMarch 23, 2010

The Problem As CMOS technology shrinks, random

defects increase

Traditionally, these defects handled with redundant rows, columns, and words to replace defective ones

As random defects increase, traditional defect strategy may no longer be sufficient

The Solution Extend the role of Error-Correcting Codes

to compensate for defects

Error-Correcting Codes (ECC) also used to compensate for transient soft errors

Find a method that allows ECCs to be used for both defects and soft errors

Multi-bit ECC Multi-bit ECC – ECC that can correct

multiple errors in one codeword

Suffers larger latency and higher coding redundancy than single error correction

Therefore unusable in L1 cache without suffering major performance issues

Overall Goal Implement multi-bit ECC in L2 cache

design to correct L2 cache defects without causing significant IPC degradation, area use, or energy cost

Steps to Success 1. Apply multi-bit ECC only to cache

blocks that require it

2. Implement buffers to limit repeated use of multi-bit ECC

3. Ensure data integrity for soft errors where ECC can no longer alone compensate for it

Limited multi-bit ECC Cache blocks with one or more defective

cells identified during memory testing; Multi-bit ECC selectively applied then

Content-Addressable Memory (CAM) then used to identify blocks requiring multi-bit ECC (referred to as m-blocks)

ISSUE: CAM requires large energy consumption

Proposed Architecture Standard L2 cache core protecting all subblocks

with single error correction, double error detection (SEC-DED) codes

Multi-bit ECC core using fully associative multi-bit ECC cache (M-ECC cache), ECC encoder/decoder, and two buffers. M-ECC cache contains location tags and corresponding check bits

Dirty Replication Cache to ensure soft error tolerance

Proposed Architecture

Multi-bit ECC Core In case of write, subblock data encoded and

check bits stored

In case of read, check bits fetched and decoded

ISSUE: Constant use of multi-bit ECC will increase latency and energy consumption at higher defect densities

Solution: Two additional buffers

Multi-bit ECC Core Buffers Pre-decoding Buffer: Small cache that keeps

copies of mostly recently accessed m-blocks; Searched before accessing M-ECC cache

Employs least recently used (LRU) policy for replacement when full; Successful due to cache access temporal locality

Reduces large amount of ECC decoding and some M-ECC cache access

Multi-bit ECC Core Buffers FLU buffer – small CAM that keeps

addresses of recently accessed cache blocks that are NOT m-blocks

Also employs LRU policy

Further reduces M-ECC cache access

M-ECC core Flow Chart

Soft Error Tolerance ISSUE: When ECC devoted to defect tolerance,

defective subblock is vulnerable to soft errors

Only necessary for blocks containing defects (including blocks with single defects protected by SEC-DED rather than multi-bit ECC)

Further, only necessary when cache block is dirty; Clean blocks can redirect to memory when soft error detected

Dirty Replication Cache Use of Dirty Replication (DR) cache

When cache block made dirty, data is also kept in this cache

When data leaves this cache, a write is performed to main memory

Ensures a backup is always available

Evaluation Cache defect density set at 0.5% Multi-ECC: BCH-based DEC-TED code (double error

correction, triple error detection); Subblocks with more than two errors repaired by redundancy

Cache subblocks contain 64 bits BCH DEC-TED decoder has parallelism of 2, uses

PGZ decoding algorithm- resulting latency of 82 cycles

Cacti 5 used to model caches; Through verilog, determined extra logic is 0.2% of area of L2 cache core

Evaluation Compared on four bases:

Base: Defect-free L2 cache with no defect tolerant functions

M-ECC only; No buffers M-ECC-pbuf: Use of predecoding buffer M-ECC-pfbuf: Use of predecoding and FLU buffers

First, determine best size of buffers for use; Then compare performance of IPC and power consumption

Size of precoding Buffer

Size of FLU buffer

Normalize IPC comparison

Normalized Power Consumption

Results Similar IPC performance, M-ECC core

power performance 30% of L2 cache core, which itself is about 10% of the entire system cache

DR Write-back hit rates L2 cache fixed at 1 MB 8-way

associative, DR varies

DR Write-back hit rates DR fully associative with 64 blocks, 1 MB

L2 cache varies

Conclusions Goal was to effectively use multi-bit ECC for L2 cache

defect tolerance at minimal performance and implementation cost

Multi-bit ECC implemented only where more than one defect found

Two small buffers included to reduce performance impact of multi-bit ECC

Dirty Replication Cache included to ensure soft error tolerance

Conclusions IPC performance nearly the same as

defect-free cache

M-ECC cache has less than 2.5% of area overhead and 36% of energy consumption overhead

Dirty replication cache has area overhead of only 0.3%, storing 96.4% of write-back data from L1 cache