lab seminar a multi-channel architecture for high-performance
TRANSCRIPT
Lab Seminar
A Multi-channel architecture for high-performance NAND flash-based storage system
Jeong-Uk Kang et.al. (KAIST)Journal of System Architecture 2007,Vol 53
Speaker : Jaewoo Lee
Flash Memory : Basic
• NOR : code storage• Random access capability
• High read performance
• Low write and erase performance
• Expensive
• Execute-in-place (XIP)
• NAND : data storage• Higher write and erase performance
• Low cost
Flash Memory : New Technology
• Multi-Level Cell (MLC) technology• Multiply the capacity of flash memory
• Decrease the operation speed high performance is challenging area
• Large Block ( compared to small block )• # of page in each block : 32 64
• Each page consist of Main area : 512 byte 2048 byte Spare area : 16 byte 64 byte
• Sequential programming restriction is added
• Multi-Channel Architecture• Each channel can operate independently
Objective
• High Performance NAND flash-based storage system• In Multi-Channel Hardware Architecture• three optimization : exploit I/O Parallelism
Striping Interleaving Pipelining
• Result Striping, Interleaving, Pipelining improve
performance up to 164%, 62%, 197% Combining all the optimization results 3.6
times improvement
Multi-channel Architecture
• Prototype System based on Multi-Channel Architecture
Multi-channel Architecture
• Channel Manager
Timing of Read and Write Op
• Read Op
• Write Op
Software Architecture
• Optimization which exploit I/O parallelism • a: striping, b: Interleaving, c: pipelining
Software Architecture : Striping
Software Architecture : Interleaving
Software Architecture : Pipelining
Back
Software Architecture : All together
Result : Impact of Optimization
• Striping
• Interleaving
Result : Impact of Optimization
• Pipeling
Result : All Together
• Throughput
• Average Service TimeS2:I2
Request Size4K
Result : Overall Performance
• Throughput delivered to User-level Applications• Read : 80% of raw , due to programmed I/O
• Write : Sync – Async -
Link
Conclusion
• This paper present• High-performance NAND flash-based storage• Exploit I/O parallelism from multiple channel• Applied three optimization
Striping Pipelining Interleaving
• We find out• Size of sub-request >= page size of NAND for striping• Request size is not related to the throughput of read
and write op for interleaving
• Result • 3.6 times improvement to conventional single
channel architecture