Paper Accepted at the Outstanding International Conference The 53rd International Conference on Parallel Processing (ICPP) 2024

(First author: Park Jun-hyuk; corresponding author: Professor Kim Young-jae)

▶ (From left) Park Jun-hyuk, master's student; Lee Chang-gyu, doctoral student; Hwang Sun, doctoral student

 

The paper 'BandSlim: A Novel Bandwidth and Space-Efficient KV-SSD with an Escape-from-Block Approach', written by Park Jun-hyuk (first author), Lee Chang-gyu (doctoral) and Hwang Sun (doctoral) of the Data-Centric Computing and Systems Laboratory with Professor Kim Young-jae (corresponding author), has been accepted for publication at The 53rd International Conference on Parallel Processing (ICPP '24).

The paper points out that the Key-Value Solid State Drive (KV-SSD), one of the storage devices attracting attention as a future technology, is still not free of the traditional block-based interface when operating under Non-Volatile Memory Express (NVMe), the high-performance storage protocol built on the PCIe interconnect. To resolve the resulting PCIe traffic amplification during data transfer and NAND page write I/O amplification during data storage, it proposes BandSlim, a new fine-grained data transfer and storage technique.

 

Noting that key-value stores in production typically store mostly small values, BandSlim proposes a fine-grained value transfer technique that carries the value inside the NVMe command itself — using the NVMe command rather than the conventional Physical Region Page (PRP) based block data DMA exchange — so that less traffic is generated on the interconnect. It also proposes a technique for packing values delivered into the device into the NAND page buffer in a resource-efficient, fine-grained way that takes account of the characteristics of the device's internal DMA engine. With these two techniques BandSlim achieved dramatic improvements, reducing PCIe data traffic by up to 97.9% and NAND page writes by up to 98.1%.

 

Park Jun-hyuk, the paper's first author, said: "I am delighted to have brought this research, which I began as an undergraduate researcher, to a good conclusion. From first finding and defining the problem through devising an appropriate solution and evaluating its performance, the whole process was hard but full of learning, and I am certain that it could never have led to this result without the passionate guidance, teaching and criticism of Professor Kim Young-jae and my two seniors, Lee Chang-gyu and Hwang Sun. I would like to express my deepest thanks to all three."

▲ (a) The software architecture of BandSlim; (b) one of BandSlim's core techniques,

the fine-grained value transfer technique using NVMe commands

▲ Performance evaluation of the fine-grained value transfer technique and the adaptive value transfer technique proposed as a response-time optimization

 

ICPP (International Conference on Parallel Processing) is one of the oldest parallel computing conferences in the world. Covering a wide range of topics across parallel computing — high-performance computing, parallel algorithms, distributed systems, cloud computing, storage architecture and more — it is widely recognized as a venue for sharing the latest technology trends and research results. ICPP 2024 will be held in Gotland, Sweden, on 12–15 August 2024. ICPP is also listed among the outstanding conferences in the software field of the Korean Institute of Information Scientists and Engineers and among the outstanding international conferences in computer science under the BK21 Plus programme (BK IF=2).

 

References:

Conference: https://icpp2024.org/

Paper: https://discos.sogang.ac.kr/file/2024/intl_conf/ICPP_2024_J_Park.pdf