Approximate Storage in Solid-State Memories Adrian Sampson - PowerPoint PPT Presentation

Approximate Storage in Solid-State Memories Adrian Sampson University of Washington Jacob Nelson Karin Strauss Microsoft Research & UW University of Washington Luis Ceze sa pa MICRO 2013

Compiler Runtime Vector GPU CPU Processor Accelerator

Network Disk Display Memory I/O Compute Storage

Disk Memory I/O Compute Storage

r o s f y t b n i e m j s a h c m r e c a r y t m m s g t c n a i x p z m I t f f jmeint correcting % 0 0 1 100% per % 0 8 80% s s output quality loss o % l 0 6 y t i 60% l a u q % 0 4 t u p 40% t u 4 o . 3 % 0 2 2 . 3 3 8 2 . 6 2 . 20% 7 4 0 % 1 2 . 0 Main-memory applications using failed blocks. s × e t 2 r i . w 2 g o l r o s n 0% e s 3 2.8 2.6 2.4 2.2 2 1.8 1.6 average write steps

Themes in approximate computing approx precise LO HI x ± y Interleaving: Error mitigation: Programs are both Exploit the hardware approximate & precise to minimize error

:) Phase-change memory (PCM) + Non-volatile Surpass DRAM’s scaling limits Faster than flash “Almost” as fast as DRAM

:( Phase-change memory (PCM) + Write speed Cells wear out & energy over time

Phase-change memory (PCM) :( Multi-level cells are denser 
 Write speed & energy but need more time and energy. Cells wear out over time Cells wear out and can no longer be used. over time

Phase-change memory (PCM) : ( Multi-level cells are denser 
 but need more time and energy 
 to protect against errors. Cells wear out over time and can no longer be used for precise data storage.

Phase-change memory (PCM) : ( Fast Dense

Phase-change memory (PCM) : ( Fast Dense Accurate

Approximate storage in PCM Trade off accuracy for performance in multi-level cell accesses. Use worn-out memory for approximate data instead of throwing it away.

Approximate storage in PCM 1 Trade off accuracy for performance in multi-level cell accesses. 2 Use worn-out memory for approximate data instead of throwing it away.

Approximate storage in PCM 1 Trade off accuracy for performance in multi-level cell accesses. 2 Use approximate throwing it away.

Single-level cells high 1 low 0 analog value digital value

Multi-level cells high 11 10 01 low 00 analog value digital value

Writing to multi-level cells high 11 probability 10 01 low 00 analog value digital value

Writing to multi-level cells, 
 approximately high 11 probability 10 01 low 00 analog value digital value

Speed Density Accuracy

Iterative writes high 11 target range 10 01 low 00 time

Iterative writes, approximately high 11 target range 10 01 low 00 time

Iterative writes, approximately high 11 target range 10 01 low 00 time

wider target range fewer iterations to converge faster writes (or better density at the same speed)

Encoding to minimize error in approximate MLC LO HI x ± y 1 cell, 4 bits 0 0 0 0 reliable unreliable

Encoding to minimize error in approximate MLC LO HI x ± y 4 cells, 16 bits 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 lots of errors

Encoding to minimize error in approximate MLC LO HI x ± y 4 cells, 16 bits 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 lots of errors

Write speedup for approximate MLC 2.5 2 best write speedup 1.5 1 0.5 0 mc smm sor fft lu zxing jmeint raytracer pa nn ml image mean main-memory benchmarks persistent data Writes are 1.7 × faster on average with quality loss under 10%

Approximate storage in PCM Trade off performance in accesses. Use worn-out memory for approximate data instead of throwing it away.

Failed cells are a fact of life 0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1 a good block

Failed cells are a fact of life 0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1 a (tragically) failed block

Traditional error correction 0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1 corrected data block correction bits

Correction resources are exhaustible 0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1 uncorrectable (bad) block correction bits e t a m i x o r p p a

Prioritized error correction LO HI x ± y 0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1 uncorrectable (bad) block correction bits e t a m i x o r p p a error exposed where it does the least harm

Lifetime extension with block recycling 2 normalized lifetime (writes) 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 mc smm sor fft lu zxing jmeint raytracer pa nn ml image mean main-memory benchmarks persistent data Lifetime extended by 23% on average or from about 5.2 to 6.5 years

Network Disk Display Memory I/O Compute Storage

Network Disk Display Memory I/O Compute Storage