Adap%ve policies for balancing performance and life%me of mixed SSD - - PowerPoint PPT Presentation
Adap%ve policies for balancing performance and life%me of mixed SSD arrays through workload sampling Sangwhan Moon A. L. Narasimha Reddy Texas A&M University 2 / 16 Outline Introduc%on Mixed SSD Arrays Workload distribu%on of
Adap%ve policies for balancing performance and life%me of mixed SSD arrays through workload sampling
Sangwhan Moon
Texas A&M University
2 / 16
– Mixed SSD Arrays – Workload distribu%on of mixed SSD array
– Online sampling – Adap%ve workload distribu%on
3 / 16
0.1 1 10 100 0.1 1 10 100 Cost ($/GB) Device Writes Per Day (DWPD, higher is be>er)
High-‑end SSDs Low-‑end SSDs
4 / 16
– Faster: PCIe interface – Reliable: SLC eMLC (write endurance = 100K) – Expensive per gigabyte
– Slower: Serial ATA interface – Less reliable: MLC TLC (write endurance < 30K) – Cheap per gigabyte
5 / 16
High-‑end SSDs
Low-‑end SSDs
3.mrr 1.r 2.w 5.(mrr + mww)⋅d 4.mrr read miss dirty entry evic%on read write wC = mrr + w NC ⋅CC wS = mww NS ⋅CS
min lC wC , lS wS ! " # $ % &
Lifetime =
mr,mw NC, NS CC,CS lC,lS wC,wS Writes per flash cell
Cache read/write miss rate The number of SSDs The capacity of SSD Write endurance of cache/storage
r,w
Read/write workload
6 / 16
High-‑end SSDs
Low-‑end SSDs
3.mrr 1.r 2.w 4.mrr read miss dirty entry evic%on read write
Item DescripKon SpecificaKon High-‑end SSD (SLC) Capacity 100 GB Write Endurance 100 K Low-‑end SSD (MLC) Capacity 200 GB Write Endurance 10 K Workload Read/write (MB/s) 100 / 250 Read/write cache hit rate 50% / 15% Read / write length 4KB / 64KB
wC = 0.5⋅100MB / s+ 250MB / s 1⋅100GB wC = 0.85⋅250MB / s 1⋅100GB
min 1.47years, 6.34years ! " # $ % &
Lifetime =
high-‑end low-‑end 5.(mrr + mww)⋅d
7 / 16
SSDs main storage
– Caching less results in poor performance – Caching more results in poor reliability
– Balance the performance and life%me of cache and storage at the same %me
metric : Latency over Life0me (less is be5er)
8 / 16
I/O requests whose sizes are 4KB are domina%ng 90% of workload is reference
Sta0c workload classifiers cannot distribute workload across cache and storage precisely
9 / 16
Frontend Cache
Backend Storage
2.mrr 1.hrr 6.mwwp 5.(1− p)w 3.mrpr 4.pw read write read miss bypassed writes dirty entry evic%on ProbabilisKc Caching Policy bypassed read miss 7.mr(1− p)r
p (threshold): the probability of caching data
p is more: cache wears out faster, performance enhances p is less: cache wears slower, performance degrades
10 / 16
LRU Main Storage LRU
Selec%ve Cache
90 % LRU LRU LRU . . . 0.1 0.2 0.3 0.9 1.0
Sampling Cache Sampling Cache Sampling Cache Sampling Cache Sampling Cache
. . . 1% 1% 1% 1% 1% LRU p 1.0 – p . . . Sampling Rate: 10%
Es%mate latency over life%me for each sampling cache Employ best value of p, the proximity of caching
11 / 16
– 13 enterprise applica%ons trace for a week
– Cache size / Storage size
– LRU, size-‑based (+ sampling), hotness-‑based (+ sampling), probabilis%c (+ sampling)
12 / 16 Threshold
0.1 1 10 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
Threshold latency life%me metric
Threshold
0.1 1 10 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
Threshold latency life%me metric
StaKc threshold based analysis Sampling based analysis Cache less Cache more Cache less Cache more Cache less Cache more Cache less Cache more
Hardware monitoring Web server
13 / 16
– Very effec%ve on some traces (mds, stg, web, prn, usr, proj, src1, src2) – Less effec%ve on very skewed workload (wdev, rsrch, ts, hm, prxy)
14 / 16
es%ma%on (beneficial) and less space for adap%ve cache (harmful)
15 / 16
caching to balance the performance and life%me of cache and storage.
proposed caching policy is effec%ve.
16 / 16