NAND B NAND-Based d SSDs: SSD

Performance States and P f Performance M Measurement t

Doug Rollins Micron Technology DougRollins@Micron.com

Flash Memory Summit 2011 Santa Clara, CA

Time-Variant Performance
1.2 10 1.0

Normalize ed IOPs

0.8

0.6

0.4

0.2

0.0 0 100 200 300 400 500 600 700 Time (Minutes)

Empirical data from several SSDs Drives were securely erased, then written with a fixed stimulus Amplitude A lit d and d shape h of f curves diff differ from f drive di t to drive di
Flash Memory Summit 2011 Santa Clara, CA

Common Terms
FOB state (red): The FOB state is visible at the extreme left of the plot. This state is reached when the drive has little to no user data; all the NAND cells ll are erased d and d available il bl to t receive i new data. d t As A the th drive di is written, the IOPs of this state decrease.
1.2 1.0 0.8 06 0.6 0.4 0.2 0.0 0 100 200 300 400 500 600 700
*Storage Networking Industry Association Solid State Storage Initiative Technical Working Group Performance Test Specification

Transition state (yellow yellow): Immediately following the FOB state, t t the th drive d i enters t a transition state, marked by steadily decreasing performance. Steady state (green): Drive performance is as described in the SNIA SSSI TWG PTS*, so this area is a steady state (by inspection) inspection). Note that: Different drives reach steady state at different times. One sample has yet to reach steady state.

FOB Transition state Steady state (desirable test range)

Normalize ed IOPs

Time (Minutes)

Flash Memory Summit 2011 Santa Clara, CA

Stimulus Sequence-Variant Performance

Example drive stimulated with: 128K sequential write (into steady state)
(followed by by) )

4K random write (into steady state)

(followed by)

128K sequential write (into steady state) Result: This drive never recovers original performance level Example drive stimulated with: 4K random write (into steady state)
(followed by)

128K sequential write (into steady state)

(followed by)

4K random write (into steady state) Result: This drive does recover original performance level
Flash Memory Summit 2011 Santa Clara, CA

Key Care-Abouts and p Enterprise p Assumptions:

Enterprise assumptions:
Drive is always full Drive is always being accessed Decisions are made on steady state performance Steady state full drive worst case

Steady state defined (from SNIA PTS):

Make sure the difference between the maximum and minimum performance is close to the average (no wide variance) AND Maximum performance (linear curve fit of the data with the measurement window) minimum performance (linear curve fit of the data with the measurement window) is small (within 10% of the average)

Full drive defined: Drive has been written over some multiple (could be 1X) of the
user-accessible LBA space by a fixed pattern that may be invariant from the test stimulus (i.e. 2X user LBA space written with 128K SEQ WRI)

Worst case defined: Drive has been stimulated over some fixed time with a
workload intentionally designed to demonstrate the drives worst possible performance. Typically this includes:
Small transfers mixed with large transfers Intentionally misaligned writes
Flash Memory Summit 2011 Santa Clara, CA

Consistent Results
Always start from a repeatable, known, fixed point:
Examples: purge, secure erase, low-level format

Always precondition the drive in the same way:

Workload-independent 128K SEQ WRI over 2X user capacity, page-aligned Workload Workload-dependent dependent apply the test stimulus iteratively until steady state is reached

Always stimulate the drive with a single, fixed stimulus until steady state is reached:
Write history affects performance Stimuli application order can affect performance Plot results Determine steady state via PTS definition

Example test sequence for every stimulus of interest:

Purge, SE, or LLF PC Stimulate until steady state is reached
Flash Memory Summit 2011 Santa Clara, CA

Example: p SLC SATA SSD

Raw IOMeter Data
SLC SATA 4K RND WRI Write Saturation
40000 35000 30000 IOPs s 25000 20000 15000 10000 5000 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Iteration Number (5-minute intervals) Iteration Transfer Number Size 1 4096 2 4096 3 4096 4 4096 5 4096 6 4096 7 4096 8 4096 9 4096 10 4096 11 4096 12 4096 13 4096 14 4096 15 4096 16 4096 17 4096 18 4096 19 4096 20 4096 096 21 4096 22 4096 % Read 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 % Random 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 00 100 IOPs 35460.98405 17382.48189 20179.53067 19460.9479 19568.64219 19657.92814 19573.7974 19561.21846 19572.46237 19597 42278 19597.42278 19577.95913 19595.8534 19442.13415 18986.06345 19065.45231 18993.24672 19016.87946 19011.96523 19024.61239 19045.18973 18990.85739 8990 85 39 18997.49995

FOB region

Transition region

Steady state region i

Flash Memory Summit 2011 Santa Clara, CA

Summary y
Assumptions and working practices:
Assume drive is always y full and always y under maximum load Assume steady state is a region of interest Always start from a repeatable, known, fixed point Always precondition each drive in the same way Always stimulate the drive with a single single, fixed stimulus until steady state is reached 4K R/W 100% random 8K R/W 100% random 128K R/W 100% sequential 8K 67R/33W 100% random OLTP 64K R/W 100% sequential

Examples of stimuli of interest (all page-aligned):

All are measured full-span Ensure that the host does not impede performance (when making relative l ti comparisons) i )
Flash Memory Summit 2011 Santa Clara, CA

Comments?

Revisit Micron Microns s FMS 2011 presentations at: www.micron.com/fms

2011 Micron Technology, gy, Inc. All rights g reserved. Products are warranted only y to meet Microns p production data sheet specifications. p Information, ,p products, , and/or / specifications are subject to change without notice. All information is provided on an AS IS basis without warranties of any kind. Dates are estimates only. Drawings are not to scale. Micron , RealSSD, and the Micron logo are trademarks of Micron Technology, Inc. All other trademarks are the property of their respective owners.

Flash Memory Summit 2011 Santa Clara, CA