TL;DR
As DRAMs shrink, they become more susceptible to multiple bit errors that current error correction methods can miss, leading to silent data corruption. This paper introduces SSCMSD, a novel scheme that combines Reed-Solomon ECC with a hash function to correct single-symbol errors and detect multi-symbol errors, significantly improving DRAM reliability without increasing read latency.
What Problem Does This Paper Address?
Modern DRAM systems are facing increasing reliability challenges due to shrinking feature sizes and higher densities. These advancements, while boosting performance, make DRAM cells more susceptible to various faults, including "disturbance errors" and transmission issues on data and address buses. Existing server-grade error correction schemes, like CHIPKILL, are primarily designed to handle one or two symbol errors per DRAM beat. However, they often fail to detect or correct more complex multi-symbol errors that can arise from faults affecting multiple data buses or DRAM chips simultaneously. This inability to detect such errors can lead to Silent Data Corruptions (SDCs), where incorrect data is silently passed to the processor, potentially causing system crashes or incorrect computations. The paper aims to address this gap by proposing a more robust error handling mechanism that can effectively detect and mitigate these multi-symbol errors.
Methods
The core of the proposed solution is the Single Symbol Correction Multiple Symbol Detection (SSCMSD) scheme. It builds upon existing error correction capabilities by integrating a hash function with a Reed-Solomon (RS) Error Correcting Code (ECC).
Here's a breakdown of the approach:
- Reed-Solomon ECC (SSC-RS): The paper utilizes a Single Symbol Correcting (SSC) Reed-Solomon code. This is a standard technique for error correction, capable of correcting a certain number of symbol errors within a data block. The specific implementation uses an RS (19, 17, 8) code, meaning it encodes 17 data symbols into a 19-symbol codeword, with 2 check symbols for error correction.
- Hash Function: A non-cryptographic hash function (specifically, a 32-bit CRC-based hash) is employed. This hash function takes the data (and potentially address information) and generates a compact "signature" or hash value.
- Combined Operation (WRITE): During a write operation, the original data is first hashed. This hash value is then combined with the data and encoded together using the RS-ECC. This ensures that both the data and its integrity check (the hash) are protected by the ECC.
- Combined Operation (READ): Upon reading data, the SSC-RS decoder first attempts to correct any single-symbol errors. If the decoder flags an error or a correctable error, the retrieved data is then passed through the hash function again. This newly computed hash is compared with the hash that was retrieved along with the data.
- If the hashes match, it's highly probable that the data is correct, and the system proceeds with minimal latency.
- If the hashes do not match, it indicates that either the data or the hash (or both) were corrupted by multi-symbol errors that the SSC-RS code couldn't fully correct or detect. In this case, SSCMSD flags an error, preventing SDC.
The paper also explores different ways to combine the hash and ECC during writes (Schemes A, B, C), ultimately favoring Scheme A where the hash is computed before ECC encoding, ensuring the hash itself is also protected by the ECC.
Datasets
Not discussed in the paper. The evaluation relies on a custom simulator that injects various fault models into encoded data.
Key Insights and Ideas
- Hash as a Second Layer of Defense: The key innovation is using a hash function not just for data integrity but as a secondary detector for errors that slip past the primary ECC. This is particularly effective for multi-symbol errors that can lead to miscorrections by the ECC decoder.
- Minimizing Read Latency: A significant design goal was to avoid increasing read latency. By performing syndrome computation (for ECC) and hash calculation in parallel, and by prioritizing the "no error" scenario, the scheme aims to achieve near-baseline read performance for error-free cases.
- Addressing Multiple Fault Types: The error model considered a range of faults, including single-bit flips, pin faults, chip failures, bus faults, and combinations thereof. This comprehensive approach ensures the scheme's robustness against realistic failure scenarios.
- Hash Aliasing is a Trade-off: While the hash significantly reduces SDCs, the paper acknowledges that hash aliasing (different data producing the same hash) is a potential, albeit low-probability, source of false negatives. The choice of a 32-bit hash aims to balance this risk with computational overhead.
Key Results
- SDC Reduction: SSCMSD effectively eliminates Silent Data Corruptions (SDCs) in the presence of multiple symbol errors, a significant improvement over the baseline SSC-RS and even extended versions of other schemes like Bamboo-ECC. Table 5 shows SSCMSD achieving 0% SDCs across various multi-symbol fault scenarios where other methods still exhibit SDCs.
- Negligible Read Latency Impact: The proposed scheme introduces minimal additional latency during read operations, especially in the common error-free scenario. The paper estimates the additional latency to be around one memory clock cycle.
- Storage Overhead: SSCMSD requires a modest increase in storage overhead. It uses 19 chips per rank compared to the baseline's 18, representing an 18.75% increase. This is attributed to the additional ECC symbols and the storage for the hash.
- Address Protection: The scheme can also be extended to protect address bits, further enhancing overall system reliability.
Limitations and Future Work
- Hash Aliasing: The primary limitation acknowledged is the possibility of hash aliasing, which could theoretically lead to undetected errors (false negatives). The paper quantifies this risk and suggests that for a 32-bit hash, the probability is very low.
- Hardware Overhead: While modest, the scheme does require additional hardware logic at the memory controller for hash computation and storage for the hash bits.
- Future Work: The paper suggests that the reliability of future memory systems can be further improved by combining SSCMSD with other protection mechanisms like AIECC, which focuses on protecting command, address, and control signals. The authors also mention the potential for using different hash functions and exploring their impact on performance and reliability.
Summary updated