RAID Cache Benefit for Avago 12Gb/s SAS

RAID Cache Benefit for Avago 12Gb/s SAS
RAID Cache Benefit for Avago® 12Gb/s SAS
MegaRAID® Controllers
White Paper
March 2015
DB10-000024-00
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Corporate Headquarters
Email
Website
San Jose, CA
globalsupport.pdl@avagotech.com
www.lsi.com
Avago, Avago Technologies, the A logo, LSI, Storage by LSI, MegaRAID, CacheCade, and CacheVault are trademarks of
Avago Technologies in the United States and other countries. All other brand and product names may be trademarks
of their respective companies.
Data subject to change. Copyright © 2015 Avago Technologies. All Rights Reserved.
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Table of Contents
Table of Contents
RAID Cache Benefit for Avago® 12Gb/s SAS MegaRAID® Controllers White Paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1 Architectural Benefits of RAID Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Performance Tests and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Streaming Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Transaction-Oriented Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Common Enterprise Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Microsoft Exchange Server 2013 Benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 SQL Server OLTP Benchmarking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Avago Technologies
-3-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Introduction
RAID Cache Benefit for Avago® 12Gb/s SAS MegaRAID® Controllers
White Paper
1
Introduction
Avago® 12Gb/s SAS MegaRAID® controllers, featuring the dual-core SAS3108 RAID-on-Chip (RoC) processor, offer
significant performance enhancements for solutions architected with 12Gb/s SAS or 6Gb/s SAS drives. The superior
read/write performance suits the controllers for a broad range of application workloads:



Enterprise data center applications, such as email, file servers, transactional databases, and analytical databases
Cloud computing, software-defined storage systems, and big data (Hadoop®)
Content applications, such as streaming video and cold storage/archival
Today’s converged storage environments contain many different storage technologies. Traditional hard disks are still
most cost effective with choices such as enterprise level 15,000, 10,000 and 7,200 RPM varieties that can be used
based on performance and cost requirements. Flash-based SAS and SATA devices add another performance spectrum
with IOPs capabilities that exceed 150,000 IOPs per device and latencies less than 35 μs. The following figure shows
the typical random write response time for different drive types when queue depth is 1.
Figure 1 Typical Latencies in the Storage Stack
SATA HDDs are well-suited for some workloads; however, they quickly reach non-linear response times when given
heavy workloads. For instance, many vendors recommend an average database latency of no greater than 20 ms to
ensure reasonable application response for end users. With only two commands, a SATA disk exceeds this
recommendation. Even the fastest 15,000-RPM disks exceed 20-ms latency with only six commands. The following
figure shows the 4-KB random writes for single 7200-RPM SATA and 15,000-RPM enterprise SAS disks.
Avago Technologies
-4-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Introduction
Architectural Benefits of RAID Cache
Figure 2 Write Latency versus Queue Depth
Latencies shown in the previous two figures are quickly becoming unacceptable for the demands of today’s storage
environment. Implementing a RAID cache product can help, as described in following sections.
1.1
Architectural Benefits of RAID Cache
Avago hardware RAID controllers use DDR SDRAM to help buffer writes to a disk and provide a fast read cache. With
intelligent cache algorithms, writes can decrease as much as 375 times when compared to a 7200-RPM SATA disk
without cache. Bursty writes can post to the RAID cache and be written back to the disk during inactive periods.
For spinning media, seek time dominates the latency for random reads and writes. If you reduce the seek distances,
such as by short stroking the drive or grouping commands to optimize seek overhead, a disk can achieve higher
performance. The Avago MegaRAID cache algorithms are highly optimized to assist with sorting, which enable the
disks to produce more IOPs. For example, an enterprise 15,000-RPM SAS HDD can achieve a maximum of
approximately 400 IOPs when given enough commands. You could easily achieve over 800 IOPs by using intelligent
RAID caching that provides optimal sort algorithms.
Additionally, when RAID cache handles write requests to the disk, the disks are free to service reads (that are not
already in cache). These cache features allow administrators to improve their disk performance. More cache means
more writes that can be handled during bursts and more requests that can be sorted. Read cache also provides
significant benefits for certain workloads, such as multiple read or write threads when adjacent requests can be
grouped to reduce disk seeks.
You can also use flash-based cache, such as MegaRAID CacheCade® Pro 2.0 software, to improve performance and
latency improvements for HDD-based RAID volumes. However, flash is susceptible to wear from repeated write and
erase cycles, whereas a RAID cache based on DRAM is not.
Avago Technologies
-5-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
Streaming Applications
It can be challenging for administrators and storage designers to understand if RAID cache will benefit their
environment. Choosing the correct RAID cache size for workload and performance requirements also presents a
challenge. Avago 12Gb/s MegaRAID SAS controllers provide many RAID-cache options, with 1-GB and 2-GB cache
sizes, and options to disable read or write cache on a per virtual disk (VD)-based policy. The next section provides
performance comparisons between these options to allow informed decisions when designing a storage
infrastructure. Keep in mind that DRAM is a volatile memory, and should be protected by using CacheVault™ or
batttery backup units designed for your specific RAID controller.
2
Performance Tests and Results
Using IOMeter 1.1, an open-source I/O workload generator, Avago designed I/O workload profiles to demonstrate
how real-world applications might perform with various cache settings and values. Sequential read and write
benchmarks reveal how streaming applications perform by simulating multiple enterprise workloads that typically
include mixed and overlapping reads and writes across request sizes, similar to as the real-world applications in the
following table.
Table 1 Real-World Applications That Benefit from Cache
Workload
Typical Applications
File and email servers Structured file systems
Enterprise databases Online transaction processing (OLTP), online analytical processing (OLAP) transactions, email applications
Multithreaded reads
Video on demand, Hadoop Framework, Hadoop Distributed File System (HDFS), cloud content,
archival backup
Multithreaded writes Hadoop source ingest (Apache Sqoop or Flume), video surveillance
Analytics
2.1
SQL, NewSQL and NoSQL databases, MapReduce
Streaming Applications
Read and write cache provides significant improvements to multithreaded sequential environments. Because
multiple threads accessing different regions can appear random to the disk, employing read ahead and write
combining techniques can greatly increase performance. The following examples use a 24-drive RAID 10 with
15,000-RPM SAS HDDs and show that cache provided up to four times higher read performance and 40 times higher
write performance simply by minimizing seek overhead.
Avago Technologies
-6-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
Transaction-Oriented Applications
Figure 3 Multithreaded Reads
Figure 4 Multithreaded Writes
2.2
Transaction-Oriented Applications
Transaction-based applications, such as databases, are very sensitive to latency to make sure SLA requirements are
met. While many databases leverage host memory as cache to speed up operations, this cache must be periodically
flushed to persistent media (via checkpoints) to reflect changes on the disk. Handling this flurry of writes with the
lowest response time helps ensure the highest database performance.
The following chart shows the latency when a burst of 16-KB random writes are sent to an eight-disk array of
enterprise class 15,000-RPM disks in five-second sequences. Without cache, the response time is four times higher,
and 2-GB cache clearly handles many more bursts than 1-GB cache while maintaining a lower average latency.
Avago Technologies
-7-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
Transaction-Oriented Applications
Figure 5 Bursty Write Performance versus Response Time (response times averaged over 10 seconds)
In addition to database requests and checkpoints, databases also use logging (or, journaling) to guarantee ACID
properties in case of failure. Maintaining high performing (and low latency) logging is a vital feature of a healthy
database. Log files tend to be small and primarily sequential, and we can imitate logging workloads with sequential
writes from 512 b to 16 KB. The following chart compares no cache, 1-GB cache, and 2-GB cache log performance
benefits. The RAID cache benefits for logging workloads has 142 times higher performance when enabling write
back cache.
Avago Technologies
-8-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
Transaction-Oriented Applications
Figure 6 Log Performance Cache Benefits
Ultimately, database performance depends on how quickly you can execute database commands such as reads,
writes, and verifies. Artificial benchmarks that imitate a 2:1 read to write ratio simulate this command sequence. The
following chart shows the cache benefit for OLTP in no cache, 1-GB cache, and 2-GB cache environments. Performance
nearly doubles with 1-GB cache, while 2-GB cache provided an additional 220 IOPs.
Figure 7 Database Transaction Performance Benefits
Avago Technologies
-9-
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
2.3
Performance Tests and Results
Common Enterprise Applications
Common Enterprise Applications
Random writes (small I/O sizes typically less than 64 KB not located adjacent to each other) are a common component
of most enterprise disk workloads including email servers, e-commerce servers, database servers, virtualized
environments, research, and analytics. Cache provides significant benefit to these workload types by improving the
IOPs and reducing the overall latency. These benefits provide a better user experience, allow higher concurrency, and
improve worker production. The following chart shows an Avago performance test with 4-KB random writes using no
cache, 1-GB cache, and 2-GB cache. With 2-GB cache, Avago measured up to 10 times more IOPs when compared to
no cache, and with 1-GB cache Avago achieved up to eight times more IOPs than no cache.
Figure 8 4-KB Random Write Performance Benefits
In addition to the IOPs benefit, implementing cache significantly lowers response times. The following chart shows
the response times collected in the same three configurations: Write Through (no cache enabled), 1-GB Write Back
cache, and 2-GB Write Back cache.
Avago Technologies
- 10 -
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
Microsoft Exchange Server 2013 Benchmarking
Figure 9 4-KB Random Write Response Time Histogram
The write back configurations handled many more I/Os during the collection period, and most completed in less than
100 μs (2-GB cache saw 15% more completions in less than 50 μs). The no cache environment completed much fewer
I/Os with response times centered around 10 ms to 30 ms.
2.4
Microsoft Exchange Server 2013 Benchmarking
The Jetstress tool, provided by Microsoft, simulates Microsoft Exchange disk I/O load on a server to verify the
performance and stability of your disk subsystem before you put a server into a production environment. The
workload consists primarily of random 16-KB read and write accesses. RAID cache still provides significant benefits by
providing low latency log writes and optimized disk writes. The Jetstress benchmark is paced based on the number of
threads, the response time, and the maximum average database access times less than 20 ms.
In this simulated environment, caching provides up to 300 times lower log and database write latencies with a 50%
increase in the number of transactions per second processed (in less than 20 ms) with 2-GB cache, and a 40% increase
in the number of transactions per second processed with 1-GB write back cache.
The following chart shows response times in no cache, 1-GB cache, and 2-GB cache environments that implement six
single-disk RAID0 1-TB 7200-RPM SATA disks. The Jetstress test used 3000 mailboxes with 200-MB per mailbox and six
Jetstress databases.
Avago Technologies
- 11 -
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
Microsoft Exchange Server 2013 Benchmarking
Figure 10 Exchange Server Database Write Response Times
The following chart shows response times in no cache, 1-GB cache, and 2-GB cache environments that implement six
single-disk RAID0 1-TB SATA disks. The Jetstress test used 3000 mailboxes with 200-MB per mailbox and six
Jetstress databases.
Avago Technologies
- 12 -
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Performance Tests and Results
SQL Server OLTP Benchmarking
Figure 11 Exchange Server Database Log Write Response Times
2.5
SQL Server OLTP Benchmarking
Using a standard SQL based application benchmark, Avago measured the transactional performance of 2,500
warehouse database scaling threads from 1 to 56 to determine RAID cache performance in a real database
environment. The number of transactions per second increase by over 300%, and increased the concurrency
capability from 8 threads to more than 50 threads.
In addition to the increase in completed transactions per second, the response time decreased by 60% with consistent
and reliable response times. Without RAID cache, response times vary significantly over the cycle of checkpoints.
The following chart shows the 2,500 warehouse database using no cache, 1-GB cache, and 2-GB cache.
Avago Technologies
- 13 -
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Summary
SQL Server OLTP Benchmarking
Figure 12 Transactions Processed per Minute versus Thread Count
The following chart shows the data response time in 10-second intervals for no cache, 1-GB cache, and 2-GB
cache environments.
Figure 13 Database Response Time
3
Summary
A primary reason most system administrator and designers use RAID is to improve reliability, availability, and capacity.
Adding RAID cache to a storage environment improves on those features with increased IOPs and consistency across
many enterprise applications. Avago benchmarking shows significant improvements in IOPs and latency by using
RAID cache for virtual disks comprised of spinning media. In addition, RAID cache can be used with flash-based cache
such as Avago CacheCade Pro 2.0 software to accelerate performance of even larger datasets.
Today’s small and medium business (SMB) and enterprise applications rely on caching to improve performance and
increase production. Battery-protected RAID cache from Avago can provide significant enhancements to today’s
Avago Technologies
- 14 -
RAID Cache Benefit for Avago 12Gb/s SAS MegaRAID Controllers White Paper
March 2015
Summary
SQL Server OLTP Benchmarking
applications by increasing overall productivity and reliability. The following table summarizes the benefit increase of
1-GB cache and 2-GB cache when compared to a no cache environment.
Table 2 Workload Cache Benefits
Workload
1-GB
Cache
Metric
2-GB
Cache
Detriment of No RAID Cache
Multithreaded Reads
MB/s
40x
40x
Disk seek overhead limits performance.
Multithreaded Writes
MB/s
400x
400x
Disk seek overhead limits performance.
Bursty Writes
Number of 16-K write
commands that a single
burst can absorb
> 2,000 > 4,000
All writes incur disk seek overhead.
Log Writes
IOPs
15x
20x
Low IOPs, unacceptable for most database
applications.
OLTP
IOPs
85%
190%
Disk-limited performance.
Small Random Writes
IOPs
8x
10x
Disk-limited performance.
Small Random Writes
Response time (μs)
30 to
100
30 to
100a
Very high response times from 10,000 ms to 30,000 ms
and dominated by disk-seek overhead.
Jetstress Database Write
Latency
Decrease in response time 250x
(ms)
250x
Low IOPS, unacceptable response times for SMB and
Enterprise Exchange Servers.
Jetstress Database Log
Latency
Decrease in response time 300x
(ms)
300x
Low IOPS, unacceptable response times for SMB and
Enterprise Exchange Servers.
Jetstress Transactions per
Second
Increase in TPS
40%
50%
Only small thread counts (less than 3 per Jetstress
database) generate acceptable database read
response times.
Transaction Database
Increase in TpmCs
Application Benchmarking
300%
300%
Scales only to 8 threads where response times throttle
workload.
Transaction Database
Decrease in response
Application Benchmarking times
60%
60%
Inconsistent response times, disks cannot handle
database checkpoints which causes an unacceptable
response time increase.
a.
15% more completions in less than 50 μs compared to 1-GB cache.
Avago Technologies
- 15 -
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising