Hyojun Kim, Sangeetha Seshadri, Clement L. Dickey, Lawrence Chiu

Phase Change Memory in Enterprise Storage Systems: Silver Bullet or Snake Oil?

IBM Almaden Research

Phase Change Memory @ engardget

2

100x faster PCM than Flash?

“Despite using the iterative process, the scientists achieved a worst-case write latency of about 10 microseconds, which represents a 100× performance increase over even the most advanced flash memory on the market today” 3

Latency vs. Throughput Paper @ Symposium VLSI Circuits, 2011

Rate: 3.2Mb/s Access time: 9.8us 32 bits (4bytes) are written during 9.8us 4 KB page write takes 200us - 1ms @ flash 4

Question #1  What is the REAL performance of PCM SSD? – [HotStorage’11] Oynx 10 GB PCM SSD • 4K random read latency: 38us • 4K random write latency: 179us

– [ADMS’12] 12 GB PCM SSD • 4K random read tency: 20us • 4K random write latency: 250us

– Both SSDs are based on 16MB chips from Micron with 90nm PCM technology • 16 bytes read time: 314ns • 64 bytes write time: 120us

 Micron provided a new PCM SSD prototype to IBM – 45nm technology, 64 GB usable capacity

 We conducted extensive performance study on this new SSD

5

Question #2  Will PCM be able to improve enterprise storage systems?  We ran tiered storage simulations with realworld storage access traces to estimate the potentials of PCM

6

Contents: Two Main Parts  Motivation  Part I: PCM SSD Performance – A prototype all PCM SSD from Micron

 Part II: Customer value estimation – PCM as storage – Performance estimation for PCM + Flash + HDD tiered storage with customer’s workload traces.

 Summary and future works 7

Part I PCM SSD Performance

8

A new PCM SSD       

9

Engineering prototype from Micron Based on 45nm PCM technology PCIe gen2 x8 interface 4KB minimum access size 64 GB usable capacity 2.6 GB/s sequential (128KB) read throughput 100-300 MB/s sequential write throughput

Comparing enterprise class eMLC Flash SSD

1.8 TB usable capacity 10

Fine-grained I/O latency measurement

Collect I/O latencies in nanoseconds 11

4KB Random Read Latency (5M samples)

PCM

6.7us

16x shorter average latency

108.0us

eMLC

12

4KB Random Write Latency (1M samples)

PCM

37.1us

3.4x longer

eMLC

13

128.3us

Summary of Performance Numbers

PCM SSD

eMLC SSD

4KB Read Latency

6.7us

108.0us

4KB Write Latency

128.3us

37.1us

 For read, PCM SSD is about 16x faster  For write, PCM SSD is about 3.4x slower – Can be improved by adding write buffer

14

Part II Customer value estimation: PCM as a Component of Multi-Tiered Storage

15

Idea PCM SSD

eMLC SSD

15K RPM HDD

4KB Read Latency

6.7us

108.0us

5ms

4KB Write Latency

128.3us

37.1us

5ms

24

6

1

Normalized Cost

 We have numbers: – PCM is 16x faster, 3.4x slower, 4x more expensive than flash

 For a given workload (traces from customers), Can we find the best combination of PCM, Flash, and HDD in terms of IOPS / $? 16

Workload Simulation 1.

1. 2.

Build a cumulative I/O distribution table from a given workload trace The portion of storage (%)

Accumulated amount of read (%)

Accumulated amount of write (%)

0

0

0

1

10

7

2

15

9

…

…

…

Assume a storage made of PCM(X%), Flash(Y%), HDD(Z%) Calculate normalized cost – e.g., Normalized Cost (50% Flash + 50% PCM) = (0.5 x 6) + (0.5 x 24) = 15

3.

Calculate expected average latency based on the cumulative I/O distribution table – Place extents on PCM based on the amount of reads as many as possible – Place extents on Flash based on the amount of reads as many as passible – Place remained extents on HDD

4. 5. 17

Calculate expected average IOPS = 1 / expected average latency Calculate IOPS/cost = expected average IOPS / normalized cost

Workload1: Retail Store

Duration

One week

Accessed capacity

16.1 TB

Amount of read

252.7 TB

Amount of write

45.0 TB

Read : Write

5.6 : 1

PCM(29%)-Flash(71%) conf. is 75% better than Flash(95%)-HDD(5%) conf. 18

Workload2: Bank

Duration

One week

Accessed capacity

15.9 TB

Amount of read

68.3 TB

Amount of write

17.5 TB

Read : Write

3.9 : 1

PCM(17%)-Flash(44%)-HDD(39%) conf. is 24% better than Flash(77%)-HDD(23%) conf. 19

Workload3: TPC-E

Duration

25.7 hours

Accessed capacity

2.8 TB

Amount of read

22.1 TB

Amount of write

4.0 TB

Read : Write

5.5 : 1

PCM(35%)-Flash(65%) conf. is 54% better than Flash(100%) conf. 20

Summary and future works

21

Summary  Performance measurement results PCM SSD

eMLC SSD

4KB Read Latency

6.7us

108.0us

4KB Write Latency

128.3us

37.1us

– PCM is somewhat slower for writes, but much faster for reads, much more predictable, and much more endurable than Flash based SSDs.

 PCM seem to be able to improve multi-tiered storage – By placing read intensive potion on PCM, IOPS/$ can be improved by 24 – 75%

22

Limitations and future works  PCM / Flash / HDD multi-tiered storage – We are missing a “Management Scheme / Algorithm” • How to find read intensive potion and migrate them to PCM with minimized write overhead? • How to let PCM absorb heavy write traffic to solve the endurance issue of MLC flash?

– PCM for application server side cache • What will happen if we use PCM instead of flash, or together with flash as an application server side cache?

 New functions on PCM – High performance key-value store based on PCM – QoS support with PCM

23

감사합니다