Optimizing Data Center Systems and Applications for Samsung SSDs

Optimizing Data
Center Systems
and Applications
for Samsung SSDs
Enabling Data Center Applications
to Leverage Samsung NVMe SSDs
Samsung Memory Solutions Lab Brochure
Optimizing Data Center Systems and Applications
for Samsung SSDs
Summary
This technology overview provides a comprehensive look at the benefits that data center applications can realize when they deploy
Samsung SSDs. The detailed technical papers referenced here demonstrate how to tune and configure high speed, data-rich applications
and storage stacks to achieve the benefits enumerated below.
Introduction
Current data center applications present significant challenges for storage sub-systems due to rapid growth in the velocity, volume
and variety of data requiring storing and processing within these applications. The latest-generation Samsung NVMe SSDs provide
unprecedented throughput and latency improvements over SATA and SAS solid state storage, while simultaneously scaling up to 10’s
of terabytes per-device. However, applications only realize a fraction of the potential improvement when older SSD and HDDs are
simply swapped for newer NVMe SSDs. By conducting extensive end-to-end performance analysis, we have identified and tuned these
applications to best leverage the performance of the latest SSDs. In some cases, tuning implies firmware, operating system alterations,
and application configuration changes. In other cases, it requires application-level code modifications.
NVMe over Fabrics (NVMe-OF) is a complementary technology that has matured this year. This technology enables network-attached
NVMe SSDs to deliver throughput and latency that is almost as good as direct-attached NVMe SSDs. NVMe-OF enables disaggregated
storage systems to deliver Samsung SSD performance in a wide variety of critical applications over most networks. Disaggregated
storage enables data centers to pool, scale and manage storage much more efficiently.
The following section summarizes our recent efforts to optimize and improve the performance of some data center applications for
direct attached and NVMe-OF scenarios.
Samsung Memory Solutions Lab Brochure
[2]
Samsung SSDs for Data Center Applications
Samsung Enterprise SSDs with V-NAND and Z-NAND Memory
Applications today need advance storage devices to process large and growing amounts of data efficiently. Different application
require different types of storage. Samsung has successfully established itself as an innovator in flash-based storage across the
enterprise markets with cutting-edge flash memory leveraging V-NAND™ and more recently Z-NAND™ technology.
Samsung’s comprehensive SSD portfolio helps to meet critical speed requirements with confidence, eradicating performance
and efficiency bottlenecks.
Samsung PM1725a
The Samsung PM1725a NVMe SSD, based on V-NAND technology,
is optimized to excel in virtually any data center scenario. This
enterprise-level, ultra-high performance SSD provides high
random read performance and is particularly suitable for readintensive data center applications. The PM1725a comes in capacity
points ranging from 800GB to 6.4TB.
Samsung PM963
Samsung’s PM963 NVMe SSD also is based on V-NAND
technology. It provides exceptional value and is more cost
effective than SATA SSDs. With approximately four times the IOPS
per GB, data center operators can continue to scale workload
utilizations on their servers, decreasing the cost of computing.
Samsung PM963 comes in the M.2 form factor and is available in
capacity points ranging from 960GB to 3.8TB.
Samsung Z-SSD™
Based on a new and highly innovative Samsung Z-NAND
technology, this new class of enterprise storage SSD’s
achieves peak performance at low queue depths, giving better
performance across all workloads. It has a unique circuit design
and a controller that can maximize performance, with 5.5 times
less latency than the Samsung PM1725a NVMe SSD. The Samsung
Z-NAND SSD achieves peak performance at low queue depths,
delivering better performance across all workloads. Customers
needing a boost in high-performance data intensive, real-time
analytics will greatly benefit from this disruptive, new technology.
Samsung Memory Solutions Lab Brochure
[3]
Samsung SSDs for Data Center Applications (continued)
Performance Specifications
SZ985
PM963
PM1725A
Data Transfer Rate (128KB data size, QD=32)
Sequential Read/Write (GB/s)
3.2 / 3.2
2.0 / 1.2
6.4 / 3.0
Data I/O Speed (4KB data size, Sustained, QD=32)
Random Read/Write (K IOPS)
750 / 170
430 / 40
1080 / 170
Latency (Sustained random workload, QD=1
Read (Typical)
20us
85us
Read (Best)
12us
N/A
Write (Typical)
16 us
50us
20us
Form Factor
HHHL
2.5”/ M.2
HHHL / 2.5”
NAND technology
Capacity
Z-NAND™
800GB
90us
V-NAND™
960GB to 3.84TB
800GB to 6.4TB
To help simplify integration of these storage devices, Samsung Multi-Stream technology is an additional feature that ensures optimal
performance and better endurance for an entire class of applications. In addition Samsung produces a wide range of SAS and SATA drives
to match the need of the application.
For more information:
Samsung Enterprise SSD:
http://www.samsung.com/semiconductor/products/flash-storage/enterprise-ssd/
Multi-Stream Technology:
http://www.samsung.com/semiconductor/global/file/insight/2016/08/Multi_stream_technology-0.pdf
Samsung Memory Solutions Lab Brochure
[4]
Applications
Samsung Enterprise SSDs with V-NAND and Z-NAND Memory
RocksDB
Key-Value stores, like RocksDB, are the basis of many data center storage system applications. The publication referenced below shows
that using Samsung NVMe SSDs provides a 3x to 4x direct benefit over Samsung’s previous generation data center-class SATA SSDs.
Samsung Multi-Stream SSDs can also provide additional performance improvements. These benefits can be achieved using PM963 class
devices. The PM1725a, and SZ985 (Z-SSD) to a greater degree, deliver even higher throughput at lower latencies.
For more information about RocksDB:
http://www.samsung.com/us/labs/pdfs/collateral/Performance-Benefits-of-Running-RocksDB-on-SSDs_Whitepaper.pdf
Cassandra
Cassandra is very a popular NoSQL database in wide-spread use by many companies, notably Netflix. We have previously shown
that NVMe SSDs can provide significant performance benefits. In a recent publication referenced below, we demonstrate Samsung
Multi-Stream technology’s performance benefits. Specifically, using cassandra-stress and a modified Cassandra version that leverages
Multi-Stream SSDs will deliver up to a 3X performance improvement. While the publication highlights SAS SSDs, multi-stream
technology is also available with PM1725 and PM983 NVMe devices.
For more information about Casandra:
http://www.samsung.com/us/labs/pdfs/collateral/Multi-stream_Cassandra_Whitepaper_Final.pdf
Throughput (OPS/sec)
10,000 •
9,000 •
8,000 •
7,000 •
6,000 •
Legacy
5,000 •
Multi-stream
4,000 •
3,000 •
2,000 •
1,000 •
0 •
100% Write
50%/50% Read/Write
Figure 1. Samsung Multi-Stream SSD Performance with Cassandra
Samsung Memory Solutions Lab Brochure
[5]
70%/30% Read/Write
Applications (continued)
ScyllaDB
ScyllaDB is an open‐source, high performance NoSQL data store, and can be considered a drop-in Cassandra replacement. However,
ScyllaDB exhibits improved performance due to a finely crafted C++ implementation that manages low-level resources including
memory, thread queues and the network stack. ScyllaDB provides a seamless migration for Cassandra applications that requires higher
performance and can nicely leverage high-performance Samsung NVMe SSDs. The referenced publication outlines PM19725-class SSD
performance benefits.
For more information about ScyllaDB:
http://www.samsung.com/us/labs/pdfs/collateral/ScyllaDB_Report_v41.pdf
Operations Per Second
10,000 •
9,000 •
8,000 •
7,000 •
6,000 •
Scylla 100%
5,000 •
Scylla 60%
4,000 •
In-memory
Cassandra
3,000 •
2,000 •
1,000 •
0 •
A
B
C
Figure 2. ScyllaDB with 2TB Data Compared to Cassandra
Samsung Memory Solutions Lab Brochure
[6]
D
Applications (continued)
Redis-on-Flash
Redis-on-Flash is a RedisLabs product that leverages NVMe SSD capacity and performance to enable customers to build applications
with memory requirements that exceed DRAM capacities. Samsung PM1725-class SSDs can deliver 2 million+ ops/sec at sub-millisecond
latencies for today’s demanding applications.
For more information about Redis-on-Flash:
https://www.slideshare.net/RedisLabs/redis-on-nvme-ssd-zvika-guz
https://redislabs.com/docs/redis-on-flash-with-samsung-nvme-benchmark/
100B Objects
2,500 •
2,250 •
2,000 •
1,750 •
KOPS
1,500 •
SATA
1,250 •
NVME
1,000 •
750 •
500 •
250 •
0 •
20%
30%
40%
50%
60%
Figure 3. RedisOnFlash Throughput with PM1725 SSD
Samsung Memory Solutions Lab Brochure
[7]
70%
80%
90%
100%
Applications (continued)
MySQL
Using TPC-C as the workload, the referenced papers show how to optimize MySQL and Linux to use Samsung PM1725 NVMe SSDs
efficiently. It shows that, when selecting SSDs, careful analysis and tuning is required to leverage all of the advantages of Samsung
NAND technology. We have shown that using PM1725-class of enterprise NVMe SSDs delivers more than 180X improvement over
traditional HDDs, using Percona MySQL Server.
For more information about MySQL:
http://www.samsung.com/us/labs/pdfs/collateral/samsung-tpcc-mysql-whitepaper-final-1.pdf
https://www.snia.org/sites/default/files/SDC/2016/presentations/solid_state_storage/VeronicaLaGrange_Accelerating_OLTP_
Performance_V6.pdf
Transactions per minute normalized to SAS-HDD 50c
250 •
225 •
200 •
175 •
180X
150 •
SAS-HDD
125 •
SAS SSD
100 •
NVMe SSD
75 •
50 •
25 •
0 •
50c
100c
150c
Figure 4. Percona MySQL with TPC-C (Percona Implementation)
Samsung Memory Solutions Lab Brochure
[8]
200c
Applications (continued)
Kafka
Kafka is a lightweight high-throughput distributed log system designed to persistently store and forward large amounts of data in the
form of records, between data sources and data processing units. Its high performance and low overhead have enabled many online
businesses from social media to shopping and music streaming to scale to the demands of increased usage. With respect to storage,
Samsung NVMe SSDs offer high performance in the throughput range of GBs per second, while also providing extremely low latency.
We can show that PM963 eliminates bottlenecks introduced by SATA devices and delivers a cost-effective solution for Kafka nodes.
Producer Troughput
6 •
•4
Log.(Speedup)
•3
4 •
3 •
•2
2 •
Speedup Factor
Records/Sec (Milliion)
5 •
•1
1 •
SATA
0 •
• 0
5
10
15
20
25
30
Figure 5. Kafka Producer Throughput with PM963 NVMe SSD
Samsung Memory Solutions Lab Brochure
[9]
35
40
45
50
NVMe
Applications (continued)
Spark
The Transaction Processing Performance Council (TPC) has recently released its first big-data benchmark called TPC-xHS V2. We have
recently completed an internal study of this workload that indicates that the Samsung PM963 data center-class device delivers a 45%
improvement over current state-of-the-art SATA SSDs (e.g. the Samsung SM863) for a scale factor of 1TB using four data nodes with
Broadwell-class Intel CPUs. Specifically, the maximum read and write throughput of the PM963 NVMe SSD is higher than that of the
SM863 SATA SSD by 84% and 78% respectively. Samsung’s data center-class NVMe SSDs (PM963) are well-suited for Hadoop clusters,
delivering superior performance when compared to any SATA device.
Disk Throughput (GB/s)
Metric (HSph@1TB)
4 •
1.5 •
3 •
1 •
2 •
0.5 •
1 •
0 •
SATA
NVMe
0 •
Read
Write
SATA SM863/863a
Figure 6. TPCx-HS Disk Bandwidth and HSph@1TB
Samsung Memory Solutions Lab Brochure
[ 10 ]
NVMe PM963
Applications (continued)
Alluxio
Alluxio, formerly Tachyon, is a memory-speed, virtual distributed storage system. It allows application compute frameworks to access
data stored in disparate storage systems at memory‐speed via a single interface. That interface is able to seamlessly access data
residing in Alluxio and any underlying file system (underFS) that it supports, such as HDFS or Amazon S3. The referenced publication
demonstrates the effectiveness of high-performance NVMe SSDs; often, a few NVMe SSDs can be cost-effectively used for performance
very close to that of DRAM. This enables Alluxio-specific storage to scale to 10s of terabytes - multiples of NVMe SSD capacities.
In the referenced publication, we describe a change made to Alluxio that enables applications to directly use the NVMe SSD at the
storage target.
For more information about Alluxio:
http://www.samsung.com/us/labs/pdfs/collateral/Alluxio-plus-NVMe-WP-v6.pdf
Throughput (OPS/sec)
50 •
45 •
40 •
40.9
38.4
35 •
GB/S
30 •
Read
25 •
Write
20 •
16.2
14.4
15 •
10 •
5 •
3.68
3.12
0 •
Memory
NVMe
Figure 7. Alluxio Throughput with NVMe SSDs
Samsung Memory Solutions Lab Brochure
[ 11 ]
SATA
Applications (continued)
VoltDB
The VoltDB has advanced the role of an OLTP database into a real-time DSS (Decision Support System). Using In-Memory architecture,
it has built a database that provides extremely high throughput for today’s real-time applications such as fraud and anomaly detection.
However, many application require high durability that volatile DRAM cannot deliver. VoltDB uses NVMe SSDs to automatically backup
and restore databases during failures while continuing to deliver much higher throughout compared to traditional databases. The
referenced paper shows that NVMe SSDs are ideal candidates if applications require a high degree of durability in an in-memory
database. The reference publication shows that the Samsung PM1725 delivers 80% more throughput, with a 40% reduction in latency,
for TPC-C like workloads.
For more information about VoltDB:
http://www.samsung.com/us/labs/pdfs/collateral/A-White-Paper-by-Samsung-Memory-Solutions-Lab-and-VoltDB.pdf
TPC-C 1,000-wh and Auto Backup
VoltDB code: version 7.2
6,000,000 •
• 25
• 20
4,000,000 •
• 15
3,000,000 •
• 10
2,000,000 •
•5
1,000,000 •
0 •
• 0
AutoBkp OFF
Sync NVMe
Figure 8. Throughput and Latency Gains with NVMe
Samsung Memory Solutions Lab Brochure
[ 12 ]
Sync SATA
Average latency (ms)
NO Transactions/Minute
5,000,000 •
Applications (continued)
NVMe-Over-Fabrics Use Cases
Data center applications with high-performance storage requirements continue to increase. Consequently, NVMe SSDs are increasingly
deployed to cater to the ever-growing demand for throughput and low-latency access. As NVMe SSDs become more prevalent in the data
center, Flash Disaggregation has been proposed to improve resource efficiency and reduce the overall cost of NVMe deployment.
NVMe-over-Fabrics (NVMe-OF) is a recent NVMe standard extension that enables remote NVMe device access over different network
fabrics. It eliminates unnecessary protocol translations (such as SCSI) along the I/O path to the remote device, exposing the NVMe
multiple-paired queue design.
Using FIO and RocksDB, the referenced article below demonstrates that NVMe-OF adds very little overhead compared to direct
attached SSDs.
For more information about NVME-Over-Fabrics:
http://www.samsung.com/us/labs/pdfs/nvmf-disaggregation-preprint.pdf
Performance (OPS)
400,000 •
300,000 •
DAS
200,000 •
NVMf
100,000 •
iSCSI
0 •
800B
10K
CPU Utilization
40 •
30 •
DAS
20 •
NVMf
10 •
iSCSI
0 •
800B Host
800B Target
10K Host
Figure 9. Throughput and CPU Utilization of RocksDB Benchmark
Samsung Memory Solutions Lab Brochure
[ 13 ]
10K Target
Applications (continued)
We recently completed testing MySQL database workloads using NVMe-OF storage devices. The graph below shows that NVMe-OF
can enable high performance NVMe SSDs to be shared across a network. This allows better NVMe SSD performance utilization, while
simultaneously enabling disaggregation that delivers additional data center management benefits.
MySQL/TPCC
3,000,000 •
• 6,000
2,500,000 •
• 5,000
2,000,000 •
• 4,000
1,500,000 •
• 3,000
1,000,000 •
• 2,000
500,000 •
• 1,000
Total IO BW (MB/s)
NO Transactions/Minute
# of MySQL/TPCC Instances
Total IO BW
0 •
• 0
0
5
10
15
20
TPC-C TpmC
25
Figure 10. Linear Scaling of MySQL w/ TPC-C using a NVMe-OF Target using NVMe SSD’s
Conclusion
The challenge of selecting the most appropriate storage device and storage architecture can be quite vexing. The wide variety of
applications and a continually evolving software ecosystem requires detailed analysis and tuning. Simply swapping NVMe SSDs for
legacy storage devices is not enough to truly realize the full benefits of the new generation of SSDs with their dramatic latency and
throughput improvements. With end-to-end performance engineering matched to the most advantageous NAND-based storage
devices, we have shown that there is a Samsung SSD for almost any data center application.
About Samsung Electronics Co., Ltd.
Samsung Electronics Co., Ltd. inspires the world and shapes the future with transformative ideas and technologies. The company is
redefining the worlds of TVs, smartphones, wearable devices, tablets, cameras, digital appliances, printers, medical equipment, network
systems, and semiconductor and LED solutions.
For the latest news, please visit the Samsung Newsroom at news.samsung.com.
Copyright 2017 Samsung Electronics, Co. Ltd. and VoltDB, Inc. All Rights Reserved. All brand, product, service names and logos are trademarks and/or registered trademarks of
their respective owners and are hereby recognized and acknowledged. Specifications and designs are subject to change without notice. All data were deemed correct at time of
creation. Samsung and VoltDB are not liable for errors or omissions. PID 2017-07
Samsung Memory Solutions Lab Brochure
[ 14 ]