Datacenter workload analysis & qualification of stroage servers fms 2017

Datacenter Workload Analysis &
Qualification of Storage Servers
Eden Kim
Calypso Systems, Inc.
Flash Memory Summit 2017
Santa Clara, CA1

Datacenter Workloads & Storage Qualification
Santa Clara, CA
2
1. Analyzing Datacenter Real World Storage Workloads (RWSWs)
2. Visualization of IO Capture Workloads
3. Reference SQL Server Workload Demo #3 at TestMyWorkload.com
4. IO Capture ‘Self-Test’ – In-Situ Application Performance
5. Creating a SQL Server RWSW Test Workload
6. RWSW Tests for Datacenter Storage
7. Comparison: (4) SATA SSDs v (1) SAS HDD v IO Capture Self-Test
8. Conclusions

3 Flash Memory Summit 2017 Santa Clara, CA.

5
An IO Stream1 is an Input/Output Operation (IO)
that has a unique:
1 Workload IO Stream definitions used here differ from SSD Endurance Streams where similar write operations are associated with a group of associated data
IO Stream Table: 2 Minute Capture Step showing IO Stream Statistics
• Random or Sequential Access
• Block Size or Data Transfer Size
A single IO Stream can occur many times during an IO Capture Step
Other Secondary Metrics can be associated with IO Streams
• Read or Write IO
• Queue Depth (QD)

6
An IO Capture is the tabulation of statistics on IO
Streams that are observed during a capture period
• IO Capture Tools gather statistics and metrics on IO Streams
• An IO Capture is NOT an IO Trace
• No data or private information is collected
• Only binary numeric tables are gathered

7
IO Capture Tools are OS specific, vary by
the OS(s) supported, capture level and IO
Metrics taken
2 IOProfiler captures and results were used for the data presented herein and can be seen as Demo No. 3 at TestMyWorkload.com
• perfmon is a public tool for Windows
• blktrace is a public tool for Linux
• hiomon by hyperIO is for Windows
• IOProfiler at TestMyWorkload.com (by Calypso) is for Windows,
MacOS, Linux & FreeBSD
SNIA SSSI is collaborating with TestMyWorkload.com offering Free Capture tools, Analytics and Reference captures for
public and SNIA Technical Works use2

8
RWSWs are key determinants in storage performance
and have a significant impact on SW and storage
optimizations
• Solid State Storage Performance depends, in large part, on RWSWs
• Unlike lab test workloads, RWSWs are comprised of dynamically changing
combinations of IO Streams & Demand Intensity
• IO Stream content affects Optimization, Design, Validation & Failure
Analysis
• IO Streams change at each layer of software abstraction in the SW Stack

9
The Software (SW) Stack refers to the layers of
software (OS, APIs, programs, drivers and
abstractions) that exist between User space and
storage.
Windows Software Stack
• IO Streams are generated in User space by software applications
• IO Streams traverse the SW Stack to storage and back
• IO Stream composition is different at different levels in the SW Stack

10
IO Streams are affected (modified) at each layer of Software abstraction.
Examples include:
• Software Applications & OS Kernel activities
• Metadata, Journaling, Caching, Packetization, Tiering
• IO Appending, Coalescing, Fragmentation, Merging
• Virtualization, Encryption, Deduplication, Compression
• Back-up, Cloning, Imaging, Snapshot, Replication
• Storage Architectures, Endurance Algorithms, SSD Optimizations
Meta-DataMerge Coalesce Append FragmentFile System Block IO VMs Cloud
Writes to cache Virtualization Encryption Data ReductionCompression De-Duplication Snapshot Back-up Clone
SANHyper Converge Scale Out Array LUN Physical NAS JBOF JBOD SSDSDS HBA

11
Solid State Storage Performance depends on how
well storage responds to constantly changing
combinations of IO Streams and Demand Intensity
RWSW – IOPS & Changing Combinations of IO Streams
RWSW - Throughput & Changing Queue Depths
• RWSWs are constantly changing combinations of IO Streams and QDs
whereas Synthetic Lab tests are a fixed and constant workload
• IO Streams will have different Block Sizes, Accesses and Read or Write IOs
• Solid State Storage responds differently to the type of access (RND or SEQ),
the Block Size, and whether the IO is a Read or a Write
• The type and combination of RWSW IO Streams and the Demand Intensity
determines, in large part, the storage performance that is provided

Santa Clara, CA
12
8. Conclusions

13
RWSW can be visualized by creating an IO
Stream Map that shows the changing
combinations of IOs and metrics over Time
IO Stream Map: 24-hour Block IO Level (Drive0) Capture showing IOPS & IO Streams
• IO Stream % probability of occurrence is shown as different color
data series and plotted against the Y-axis
• IOPS are shown by the dominant black line and are plotted against
the secondary Y-axis
• Time is shown along the X-axis: 24-hour capture at 2 minute steps
• Secondary metrics captured by the Capture tool can be displayed

14
Secondary IO Metrics captured by the
IO Capture tool can be viewed on an IO
Stream Map
• IO Capture tools can capture various Secondary IO Metrics
• IO Streams can be listed by RND/SEQ access, Block Size & R/W IO
• Average & Maximum Response Times and QDs are shown to the left
• Additional IO Metrics can be shown for IO Count, Duplication Ratio,
Compression Ratio, Disk Utilization, Reads, Writes and more

15
IO Streams can be filtered, extracted or presented
by specific Application IOs
• Here we extract only the sqlservr.exe IOs
• The IO Stream Map (top) shows sqlservr.exe IOs that occur > 3% of the
time over the course of the 24-hour IO Capture
• The Cumulative Sqlservr.exe Capture (workload box ‘C’ - lower left)
shows a total of 395 Streams & 36 separate Processes
• (6) sqlservr.exe IO Streams are 78% of the Total IO Streams
• The total sqlservr.exe application IOs are 79.9% of the Total IO
Streams

16
RWSWs let us extract specific application
IOs to analyze the IO Composition, Metrics
and Performance that occur on the target
server during the IO Capture. We can see:IO Stream Composition of Selected Processes Changing Combinations of IO Streams
Application Process Response Times Throughput and Queue Depths v Time
• IO Stream Distribution of the selected Process(es)
• Changing combinations of IO Streams over time
• IO Process Average & Maximum Response Times
• IO Process Throughput and Average & Maximum Queue Depth
• Secondary IO Metrics gathered by the IO Capture tool

Santa Clara, CA
17
7. Comparison: (4) SATA SSDs v (1) SAS HDD v Capture Self-Test
8. Conclusions

Santa Clara, CA
24
8. Conclusions

25
Self-Test is the presentation of IO Capture Metrics
gathered during the IO Capture event.
• Also known as ‘In-Situ’ Server Performance Analysis
• Creates Performance Plots based on the host server activity
• Allows analysis and optimization of SW stack
• Provides a basis to create RWSW test stimuli

27
6 Dominant IO Streams
over 24-Hour Capture
• RND 8K W 4.4%
• SEQ 64K R 5.6%
• SEQ 8K R 14.2%
• RND 8K R 16.4%
• SEQ 0.5K W 28.9%
• RND 64K R 30.5%

28
24-Hour sqlservr.exe workload
IOPS v Time
• IO Rate Throttled by Server Applications
• Relatively Low Disk Utilization
• Limited Demand Intensity
• HDD based storage

29
Periods of Differing Demand
Intensity (Queue Depths) and
Throughput
• 2 am Back-up Throughput (TP) Spike
• Low Queue Depths (QD) in Early Morning Hours
• High Queue Depths during Day time
• Varying Average QDs over 24-hour capture

30
Periods of Differing IO Stream
Combinations, Disk Utilization,
IOs and IOPS
• 2 am Back-up IOPS Spike
• Low Disk Utilization in Early Morning Hours
• High IOPS and Mixed IO Streams during Day time
• Varying IO Count over 24-hour capture

31
Periods of Differing Average
and Maximum Response
Times over 24-Hours
• Higher Response Times (RTs) with more
Users (QD) / Disk Utilization
• Lower Response Times (RTs) with fewer Users
(QD) / Disk Utilization
• Response Times reflect total server, application
and storage activities

32
Average IOPS &
Average Response Time
• Limited IO Rates from server
• Periods of Low Disk Utilization
• Periods of Low Demand Intensity
• HDD based storage

33
Compression Ratio &
Duplication Ratios
• CR & DR measures data written to storage
• CR - measures how much MORE data can be
compressed
• DR – measures how many written blocks are
duplicated
• Compare File system and Block IO level to
assess efficacy of Compression & Dedupe

Santa Clara, CA
34
8. Conclusions

35
The 24-Hour sqlservr.exe workload is used
to create two RWSW test workloads:
1. Composite 6 IO Streams – 6 IO Streams for every test step. Used for
Multi-WSAT, Individual Streams-WSAT & DIRTH tests. i.e. each test
step applies the same combination of 6 IO Streams.
2. 24-Hour Replay – reproduce each capture step combination of IO
Streams. Used for Replay-Native test. i.e. each test step reproduces
the unique combination of IO Streams that occurred in the original IO
Capture.

Santa Clara, CA
36
8. Conclusions

37
1. Multi-WSAT – applies fixed 6 IO Stream composite workload
for each test step to Steady State.
2. Individual Streams-WSAT – tests each individual IO Stream as a
separate Steady State measurement.
3. DIRTH (Demand Intensity Response time Histogram) - applies
fixed 6 IO Stream composite workload across a range of 1 to
1,024 Users to measure IOPS & Response Time saturation.
4. Replay-Native - reproduces each capture step combination of IO
Streams, Queue Depths and Idle Times for storage comparison to
the original IO Capture server storage
Multi-WSAT Individual Streams-WSAT
DIRTH Replay-Native

Santa Clara, CA
38
8. Conclusions

39
The fixed Composite 6 IO
Stream workload is applied to
Steady State
• Samsung SSD has highest performance
followed by Seagate and Micron SSDs
• Sandisk SSD and Seagate SAS HDD are lower
• See component IO Stream performance in the
following Individual Streams-WSAT results

40
Each component IO Stream is
applied to Steady State
• Samsung SSD has highest overall performance
• Sandisk SSD has highest RND 8K R
• Micron SSD has highest RND 8K W
• Note the SEQ 0.5K W performance differential
• SEQ 0.5K W are 28.9% of the total workload and
explaining the difference in overall performance

41
The Fixed IO Stream composite
workload is applied across a
range of 1 to 1,024 Users
• Host Server QD Range is Ave QD 5, Max QD 306
• Test Performance at 32 Users is highlighted (red box)
• Note that Ordinal Rank is consistent across Users

42
Reproduces the IO Capture
step IO Stream combinations,
QDs and Idle times for 24-Hrs
• Seagate SSD is second highest
• Micron SSD is close to the group mean
• Sandisk SSD and Seagate SSD are lower
• Note Self-Test reference (red dotted line)

Santa Clara, CA
43
8. Conclusions

44
• Understanding Your RWSW is critical for Datacenter, Storage
Server and SSD Design and Optimization
• Know what you are optimizing (IO Streams & SW Stack levels)
• Find out what IO Streams are presented to Storage
• Try Demos, Free Captures & Analysis at TestMyWorkload.com
• See SNIA Reference Captures at TestMyWorkload.com
• Download the White Paper at TestMyWorkload.com for full data
analysis of the presented tests and sample pool drives
• www.TestMyWorkload.com

45
www.TestMyWorkload.com
info@calypsotesters.com

Datacenter workload analysis & qualification of stroage servers fms 2017

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Datacenter workload analysis & qualification of stroage servers fms 2017

Similar to Datacenter workload analysis & qualification of stroage servers fms 2017 (20)

Recently uploaded

Recently uploaded (20)

Datacenter workload analysis & qualification of stroage servers fms 2017