Transcript
SM953 White Paper The Ultimate NVMe SSD for Data Center
© 2014 Samsung Electronics Co.
SAMSUNG ELECTRONICS RESERVES THE RIGHT TO CHANGE PRODUCTS, INFORMATION AND SPECIFICATIONS WITHOUT NOTICE. Products and specifications discussed herein are for reference purposed only. All information discussed herein is provided on an “AS IS” basis, without warranties of any kind. This document and all information discussed herein remain the sole and exclusive property of Samsung Electronics. No license of any patent, copyright, mask work, trademark or any other intellectual property right is granted by one party to the other party under this document, by implication, estoppels or otherwise. Samsung products are not intended for use in life support, critical care, medical, safety equipment, or similar applications where product failure could result in loss of life or personal or physical harm, or any military or defense application, or any governmental procurement to which special terms or provisions may apply. For updates or additional information about Samsung products, contact your nearest Samsung office. All brand names, trademarks and registered trademarks belong to their respective owners. © 2014 Samsung Electronics Co., Ltd. All rights reserved. 416, Maetan 3-dong, Yeongtong-gu, Suwon-si, Gyeonggi-do 443-772, Korea www.samsung.com 2014-09 Revision History Version 0.1 Date
Author
Approver Amendment
02
Lee Won-ju, Principal Engineer Chang Jong-baek, Principal Engineer Song Sang-hoon, Senior Engineer Kim Jae-eun, Senior Engineer Lee Sang-geol, Senior Engineer Koh Seung-wan, Senior Engineer Park Hae-sung, Engineer Kim Sung-wook, Engineer Jang You-jin, Engineer Na You-jung, Assistant Engineer Choi Young-gil, Assistant Engineer
SM953 White Paper
Contents
Introduction to SM953
04
Features of SM953
05
-
Controller: UBX NAND: 19-nm MLC Interface: PCIe 3.0 Protocol: NVMe 1.1 Host Controller: CPU Power Form Factor Hot Swap Bootability Management Component Transport Protocol (MCTP) Endurance and Warranty
Performance and Applications of SM953
-
11
Basic Performance Redundant Array of Independent Disks (RAID) Performance Server Virtualization Web Server Application Server DB Server
Conclusion
18
Appendix
19
-
Specifications Evaluating the Performance Server Application Workload Firmware Update Analysis Tool
03
Introduction to SM953
Data increases geometrically with the widespread use of IT devices, including smartphones and tablet PCs. Storage is becoming more important in processing big data quickly and providing the best service. Some years ago, the HDD was used as the main storage device. Nowadays, however, the SSD is increasingly utilized to improve the storage processing speed. Nonetheless, because of the limitation of SATA/AHCI as the Legacy of the HDD, performance is not fully maximized. This white paper describes how the Samsung SM953 is a PCIe/NVMe product that overcomes the limitation in the SATA interface speed and AHCI processing, providing three times higher performance compared to the existing SATA SSD product. It is the best datacenter-oriented storage in the big data era. The Samsung SM953 is a PCIe Gen.3 product that supports up to four lanes and two types of form factors (M.2/2.5”). It is mounted with the latest 19-nm MLC, providing high-performance sequential R/W 1,750/850 (MB/s) and random R/W 250K/16K (IOPS). The next chapter will describe the details of the Samsung SM953, the datacenter-oriented NVMe. • • • • • • • • •
04
High-performance and most advanced Samsung UBX controller technology High-endurance and reliable 19-nm MLC NAND flash PCIe Gen.3 four-lane interface and the NVMe 1.1 protocol 128KB sequential read/write: 1,750/850MB/s (@480GB) 4KB random read/write: 250K/16 KIOPS (@480GB) Low power consumption and advanced power management 480GB and 960GB (2.5-inch only) capacity M.2 and 2.5-inch form factors 0.9 DWPD (@480GB) and a 5-year warranty
SM953 White Paper
Features of the SM953
This chapter discusses the NVMe interface and the H/W and S/W features of the SM953, such as controllers and NAND. Aside from the NVMe, this chapter includes a basic description of the SATA/AHCI communication and each chipset’s (layer) features, which affect the NVMe performance, such as the CPU/PCH. In addition, it provides some features that should be considered for the NVMe SSD, including bootability and hot swapping. The key features of the SM953 are shown in Table 2-1.
Features
SM953
Process technology PHY/Link Link Power Management CMD layer NAND structure DRAM CPU Target NAND ECC Security NAND Interface
Host
System
Flash
32nm PCIe Gen3 x4 Support(optional) NVMe 8ch / 8way LPDDR3-1600 3 * Cortex-R4 @ 500MHz 19nm MLC BCH AES256 Toggle 533Mbps
[Table 2-1] SM953 Product Features
Controller : UBX The SM953 has a UBX controller, and the command set supports both the NVMe and AHCI. The SM953 uses the NVMe command set, and the key features of the NVMe are shown in Table 2-2.
Features
SM953
Command Set Multiple Namespaces Arbitration Mechanism Logical Block Size Interrupt Power Management
Admin/NVM Command Set support Up to 4 Namespaces Support 512/4K/8KB INTx/MSI/MSI-X APST(2 operational state 6W/8W) [Table 2-2] SM953 NVMe Features
As the host interface, the PCIe Gen.3 four-lane high speed is supported; therefore, it can use the NAND flash without an interface bottleneck. The NVMe is compliant with specification 1.1a, supporting up to four namespaces and one administrator, up to eight I/ O queues, and up to 64K of entries for each queue size. To realize low-power consumption, it supports the PCIe link power management. For Active State Power Management (ASPM) and
05
deep power down, it supports the L1.2 feature. The logical sector size is 512B/4KB/8KB and can be set for each namespace. For security, it is built with the AES256 engine. With regard to data integrity, it supports full data path E2E protection. Aside from the pin-base/MSI interrupt for compliance with the Legacy systems, it supports the MSI-X interrupt to provide features optimized for the multi-queue environment of the NVMe. It also provides I2C bus as the Management Component Transport Protocol (MCTP) for device management.
NAND : 19-nm MLC The NAND flash chips used in the SM953 are applied with a high-performance/quality 2-bit Multi-Level Cell (MLC) manufactured with the latest 19-nm processes. For a more improved lifecycle, the flash chips are tested in extreme conditions with the SSD components at the system level. As one of the most important error correction codes, the signal-processing algorithm applied to the SM953 detects signal inconsistency in real time and improves the errors in advance to ensure the reliability of the data read from the NAND flash chip. The other algorithms are designed to specially monitor the lifetime of all NAND flash cells and adjust the cell operation based on each cell’s conditions for more improved durability. In addition, the periodically created NAND flash chip lifetime logs are used to help the algorithms find the best solution and extend the SSD lifetime as much as possible.
Interface : PCIe 3.0 The SM953 is the PCIe SSD based on the NVMe protocol. The PCIe interface can provide 1GB/s transfer bandwidth (based on PCIe 3.0) with only one lane, offering a 1.7 times wider bandwidth than the SATA Gen.3 6GB/s (600MB/s). With the decreased bottleneck shown in the existing SAS (1,200MB/s) and SATA (600MB/s) SSD interfaces, users can experience four times the high-speed performance of the PCIe Gen.3, up to 4GB/s. PCIe Gen3 x4 lanes 4,000MB/s
2,000MB/s
PCIe Gen3 x2 lanes
1,000MB/s
PCIe Gen3(8Gbps)
PCIe Gen2(5Gbps) SATA2.0(3Gbps)
2007
2008
SATA3.0(6Gbps) Today
2009
2010
2011
2012
2013
2014
2015
[Figure 2-1] PCIe Interface for Flash Storage
For a SATA device, multiple SSDs should be connected through the RAID configuration for high-speed performance. Note, however, that the PCIe SSD provides the efficient performance of two SSDs with only one device, lowering total cost of ownership (TCO). The basic protocol stack of the PCIe SSD and SATA SSD is illustrated in Figure 2-2. The SATA interface design is HDD-based. Since communication with the host is made via a Host Bus Adapter (HBA), performance may deteriorate when a layer is added. Note, however, that the PCIe SSD lowers the interface bottleneck by connecting to the host directly, not via the HBA.
06
SM953 White Paper
Host
Host Application OS
Application HBA(Host Bus Adapter)
ATA Command Set
SATA SSD
ATA Command Set
AHCI(SATA FIS)
AHCI(SATA FIS)
PCIe Transaction
PCIe Transaction
PCIe Link PCIe PHY
ATA Command Set
PCIe SSD
OS ATA Command Set
ATA Command Set
AHCI(SATA FIS)
AHCI(SATA FIS)
PCIe Transaction
PCIe Transaction
SATA Transport
SATA Transport
PCIe Link
SATA Link
SATA Link
PCIe Link
PCIe Link
PCIe PHY
SATA PHY
SATA PHY
PCIe PHY
PCIe PHY
SATA
PCIe
PCIe
[Figure 2-2] Comparison of the SATA SSD Protocol Stack and PCIe SSD Protocol Stack
Protocol : NVMe 1.1 The existing AHCI has performance limitations because of its architectural limit; hence the difficulty in improving system performance by improving the SSD performance only. The NVMe was developed to overcome the limitations of the AHCI driver. Unlike the AHCI, which has been developed for the HDD and optimized for sequential processing, the NVMe can process 64K of commands by queue and 64K queue optimized for parallel SSD configuration. In addition, the AHCI needs four commands to access the uncached data, whereas the NVMe can process it immediately.
AHCI Uncacheable Register Read MSI-X and Interrupt Steering Parallelism & Multiple Threads Maximum Queue Depth Efficiency for 4KB Commands
NVMe 4 per command No Requireds synchronization Lock to issue command 1 Queue 32 Commands per Q Command parameters Require two serialized host DRAM fetches
0 per command Yes No locking, doorbell Register per Queue 64K Queues 64K Commands per Q Command parameters in one 64B fetch
[Table 2-3] Comparison of NVMe and AHCI
With the streamlined storage stack, the NVMe enables shorter response time than the existing SATA/AHCI and higher performance by supporting multiple queues. Therefore, it offers great performance advantages that cannot be provided by the SATA/AHCI in a heavy workload server environment.
Host Controller : CPU The best advantage of the NVMe is that it provides the highest performance by removing the HBA bottleneck by connecting directly to the CPU. The NVMe performance is affected by the CPU's core and frequency. For the highest performance, a certain number of cores and clock speeds are required. Figure 2-3 shows the evaluation results for the NVMe SSD, which provides the best performance with a four-core CPU and at least a 2.5GHz clock speed.
07
4KB Random Read
1W
2Core + 2.9GHz
IOPS
4Core + 2.0GHz
4W
8W
16W
32W
64W
4Core + 2.9GHz
800,000
MAX. Performance
700,000 600,000 500,000 400,000 300,000 200,000 100,000 0
0
2
4
8
16
32
64
128 256 QD
0
2
4
8
16
32
64
128 256 QD
※Test 0condition : refer Appendix. 2 4and workload 8 16 32 64to 128 256 QD
[Figure 2-3] NVMe Best Performance by Number of Cores and Frequency
The multi-core system generally used for a server system affects the NVMe performance considerably. In the Non-Uniform Memory Access (NUMA) structure, the soft interrupt between the CPUs may cause low performance. In addition, the processing-interrupt performance for the allocated I/O submission may vary according to the location of the PCIe slot where the SSD is connected.
Local Memory
Local Memory Remote Memory Accesses
1
HW interrupt 2
CPU 0
Soft Interrupts
CPU N
PICe Slot
HW interrupt
PICe Slot Request Queue Lock
Acquire/Release Ownership
Data I/O Flow in NUMA architecture ( 1 , 2 )
[Figure 2-4] I/O Processing in a Multi-Core System
Power On the server side, where the I/O frequently occurs, there is not much need for power management, unlike the client side. Note, however, that the SM953 supports various power levels and low-power features (L1.2) according to customer requirements. By adjusting the I/O delay time, it can set the RMS power value based on the PCI slot's power limit. In addition, it provides optimized values for the highest performance and power.
Form Factor The SM953 supports two types of form factors: the small, thin, and lightweight M.2 form factor and the hot-pluggable 2.5-inch SFF8639. The proper form factor is selected according to the application. Both types support the data path up to the PCIe four-lane interface.
08
SM953 White Paper
Controller Controller PMIC PMIC NAND NAND
Tantal Capacitor Tantal Capacitor NAND
DRAM DRAM
< SFF-8639 2.5" Form Factor >
< M.2 Form Factor (110x22mm) >
[Figure 2-5] SM953 Form Factor Types
Hot Swap Hot swap involves replacing the SSD, HDD, CD-ROM drive, power supply or other devices while the computer system is running, requiring no shutdown/rebooting. A device is replaced when it fails or when the data needs to be replaced with other data. Hot plug is a term similar to hot swap, which means that the operating system normally recognizes the device even when the device is removed (hot removal) or inserted (hot add) while the system is running. For the SM953, the SFF-8639 form factor SSD supports the hot plug function; multiple tantalum capacitors ensure stable data integrity even when the system is in sudden power off recovery (SPOR) state. Currently, however, the M.2 form factor SSD does not support this function.
Bootability Bootability covers all procedures from installing to booting an OS in a storage device. For these procedures, the storage device should be recognized as a bootable device on the system because many compatibility issues can arise. For example, when a Legacy BIOS is used, the BIOS should recognize the storage device, which should be registered to the BIOS to include the storage device in the booting list. For the existing SAS/SATA-type of storage, the storage is recognized by a Platform Controller Hub (PCH) chipset and included in the booting list. When a HBA or RAID card is used, the ROM option in the HBA and RAID card is used to notify the BIOS (Legacy) or UEFI that it is a device on which an OS can be installed and booted. An additional booting device can then be set in the HBA or RAID card. Nowadays, since various vendors request the NVMe SSD booting function, the SM953 selectively provides the ROM option. However, it does not provide the ROM option because of possible compatibility issues with the BIOS and chipset.
Management Component Transport Protocol (MCTP) The SM953 provides the SMBUS-based MCTP. The SMBUS is a two-line bus based on the I2C serial bus protocol that consists of a clock and data command. If the PCIe link up is not available, regardless of the in-band state, it is used to check the SSD state, such as Smart Log or the device ID, or have a host check the temperature and device state frequently. For the SM953, only the 2.5-inch SFF8639 form factor supports this function.
09
Endurance and Warranty The SM953 guarantees a reliable and stable lifetime. Its specifications guarantee a 0.9 Drive-Write-Per-Day (DWPD) endurance and 1-month data retention with a five-year warranty. All lifetime evaluations are based on the Joint Electron Device Engineering Council (JEDEC).
RELIABILITY SPECIFICATIONS • Uncorrectable Bit Error Rate • MTBF
2,000,000 hours
• Power on Cycles (Ambient)
50,000
• Component Design Life
5 years
• Endurance - 480GB • TBW(@4KB Random Write) - 480GB • Data Retention
10
1 sector per 1017 bits read
0.9 DWPD 750 TB 1 months
SM953 White Paper
Performance and Applications of the SM953 This chapter describes the basic performance of the SM953 and the RAID performance. In addition, it discusses the datacenter applications that can maximize the effect of deploying the SM953 and the performance enhancements the SM953 provides compared to an existing SATA SSD. The performance results are from the SM953 M.2 480 GB SSD.
Basic Performance Sequential read/write performance and random read/write performance represent the basic indexes that can show excellence of performance in an actual application. As a result of comparing the SM953 product to the latest SATA 6Gb/s SSD, the SM953 exhibits about four times higher performance than the SATA SSD, as shown in Figure 3-1. Figure 3-2 presents the IOPS consistency comparison, which indicates how consistent the performance can be. Since the SM953 provides consistent and high performance with slight changes compared to the SATA SSD, it is suitable for the datacenter that requires high, consistent performance.
SATA 480GB
SM953 480GB
2,000
SATA 480GB
SM953 480GB
300,000
Single Performance(Random)
200,000
1,760
1,600
Single Performance(Sequential)
1,200
180,000
800
120,000
400
60,000
0
0 128KB Sequential Read
4KB Random Read
128KB Sequential Write
4KB Random Write
※ Refer to the Appendix for test conditions and workload. [Figure 3-1] Basic Performance Comparison
SATA 480GB 300,000
SM953 480GB
SATA 480GB
SM953 480GB
18,000
IOPS Consistency=99.4%
stdev=172
17,000
250,000
16,000
200,000
15,000 150,000 100,000 50,000
14,000
IOPS Consistency=99.3% IOPS Consistency(4KB Random Read)
13,000 12,000 Total Time : 30min
stdev=342 IOPS Consistency(4KB Random Write)
Total Time : 30min
※ Refer to the Appendix for test conditions and workload. [Figure 3-2] Comparison of IOPS Consistency Features
11
RAID(Redundant Array of Independent Disks) In a datacenter, the RAID environment is generally configured to increase performance or safety. RAID0 (striping), RAID1 (mirroring), and RAID5 (distribute parity) are the general configurations. Figure 3-3, Figure 3-4, and Figure 3-5 compare the performance of a combination of the SATA SSD RAID0, RAID1, and RAID5, single SM953 performance, and RAID performance, respectively. In most cases, the single SM953 performance is superior to the performance of the SATA RAID. If the SM953 is configured as RAID, the performance difference will be greater, proving the high-performance effect of the SM953. SATA SSD(2Disk RAID)
SM953 480GB(2Disk RAID)
SM953(Single)
SM953 480GB(2Disk RAID)
S/W RAID0(Sequential)
SM953(Single)
S/W RAID0(Random)
0
16,537
120,000
263,153
888
1,536
240,000
769
908
1,400
111,821
360,000
1,760
2,100
523,720
480,000
13,184
3,184
2,800
700
SATA SSD(2Disk RAID) 600,000
32,976
3,500
0 128KB Sequential Read
128KB Sequential Write
4KB Random Read
4KB Random Write
※ Refer to the Appendix for test conditions and workload. [Figure 3-3] RAID0 Performance Comparison
742
120,000
417
448
400 0
S/W RAID1(Random)
60,000
16,558
180,000
800
SM953 (2Disk RAID)
262,363
1,200
SATA SSD(2Disk RAID) 300,000 240,000
1,717
1,600
S/W RAID1(Sequential)
6,447
SM953 (2Disk RAID)
65,276
SATA SSD(2Disk RAID) 2,000
0 128KB Sequential Read
128KB Sequential Write
4KB Random Read
4KB Random Write
※ Refer to the Appendix for test conditions and workload. [Figure 3-4] RAID1 Performance Comparison
SATA SSD(3Disk RAID)
2,400
400,000
2,568
450,000
300,000 200,000
226
800 400
522
250,000
1,200
1,465
1,600
0
S/W RAID5(Random)
392,080
350,000
187,773
2,000
SM953(3Disk RAID)
150,000 100,000 50,000
22,459
S/W RAID5(Sequential)
18,785
SM953 (3Disk RAID)
SATA SSD(3Disk RAID) 2,800
0 128KB Sequential Read
128KB Sequential Write
4KB Random Read
4KB Random Write
※ Refer to the Appendix for test conditions and workload. [Figure 3-5] RAID5 Performance Comparison
12
SM953 White Paper
Server Virtualization In the server virtualization environment, shown in Figure 3-6, multiple virtual machines share the same storage resource. In this environment, multiple users concurrently access the storage and create a huge workload. Therefore, a product suitable for multicommand processing can deliver great performance. Figure 3-7 shows the performance comparison of the SM953 and SATA SSD in a virtualized environment of Microsoft® Hyper-V®. As the number of virtual machines increases, the SM953 exhibits higher performance because of its higher multi-queue utilization. Therefore, deploying the SM953 in the virtualized environment offers a greater advantage. VM1
VM2
VM3
APP
APP
APP
Guest O/S
Guest O/S
Guest O/S
Hypervisor
[Figure 3-6] Server Virtualization Environment SATA SSD
Virtual Machine Performance [KIOPS]
SM953
300
250
200
59.5 58
58
106
57.6
50
171
150 100
262
250
0 VM x 1
VM x 2
VM x 4
VM x 8
※ Refer to the Appendix for test conditions and workload. [Figure 3-7] Performance Comparison by Number of VMs
The general performance of the SM953 has been described above. The next section will discuss the effect of deploying the SM953 on general datacenter servers, web servers, application servers and database servers.
Web Server The web server forwards web pages to the client via the HTTP. Popular servers are the Apache™, Internet information server (IIS) and enterprise server. Nowadays, the web acceleration server is used to improve the response speed of a web server and to reduce the load; it caches the contents and compresses and transfers data. Therefore, storage device performance plays an important role. Figure 3-9 compares the performance of the SM953 and SATA SSD in the web server with a moderate workload level. The SM953 shows two or three times higher performance compared to the SATA SSD.
13
Users
Web App Severs Databases Network Firewall
Web Accelerator
Web
[Figure 3-8] Web Acceleration Server
Web Server Application Performance
SM953
24,000
8,000
4,332
16,000
0 Web Servers
12,961
36,454
32,000
17,740
Performance (IOPS)
SATA SSD 40,000
Media Streaming
※ Refer to the Appendix for test conditions and workload. [Figure 3-9] Performance Comparison with Web Server Workload
Application Server The application server interworks with the database and processes the user’s dynamic server contents. The performance of the mail server – one of the application servers accessing the storage device frequently – largely depends on the storage performance. Figure 3-10 compares the SM953 and SATA SSD in the application server with a moderate workload level. The SM953 shows two times higher performance compared to the SATA SSD. Figure 3-11 and Figure 3-12 illustrate the results tested with Jetstress 2013, a tool that evaluates the server performance of the Microsoft Exchange (one of the mail servers). The result show that the SM953 has 1.1 times higher Achieved Transactional I/O per Second (TPSE) performance and 2.9-6.3 times higher latency property compared to the SATA SSD.
14
SM953 White Paper
Mail Server Application Performance
SM953
32,000
16,000 8,000 0
21,444
24,000
10,426
Performance (IOPS)
SATA SSD 40,000
Exchange Email
※ Refer to the Appendix for test conditions and workload. [Figure 3-10] Performance Comparison with Mail Server Workload
6,000 4,500
6,399
TPSE(Achieved Transactional I/O per Second)
SM953
5,799
Performance (IOPS)
SATA SSD 7,500
3,000 1,500 0
※ Refer to the Appendix for test conditions and workload. [Figure 3-11] TPSE Comparison with Jetstress 2013
DB Read Latency (ms)
SATA SSD
4.8
4.0
5.6
5.0
3.0
2.4
2.0
1.2 0
0.9
3.6
DB Write Latency (ms)
SM953
1.0
1.5
SM953
4.4
SATA SSD 6.0
0
※ Refer to the Appendix for test conditions and workload. [Figure 3-12] Latency Comparison with Jetstress 2013
Database Server The database server saves and processes integrated information from several servers as a common data bundle. "Data is saved and accessed frequently in a general database server environment. Therefore, storage performance is extremely important. Figure 3-13 compares the SM953 and SATA SSD with a moderate workload to the general database server storage. The SM953 exhibits 2.6 times higher performance than the SATA SSD. Figure 3-14 presents the TPC-C results, which models the transactions in the online e-commerce process; a database server receives commands from several devices connected online, updates the data in the database, and then returns the processing results to the connected devices. The SM953 shows 1.1 times higher I/O Transactions
15
Per Second (TPS) compared to the SATA SSD. That means that the SM953 has about a 10 percent faster transaction processing capability on the e-commerce application compared to the SATA SSD. DB Server Application Performance
SM953
21,459
20,000
15,990
15,000 10,000 5,000
7,624
8,088
Performance (IOPS)
SATA SSD 25,000
0 Database OLTP
Decision Support System
※ Refer to the Appendix for test conditions and workload. [Figure 3-13] Performance Comparison with the Database Server Application Workload
The Not Only SQL (NoSQL) database is optimized to process unformatted data, such as documents and images that cannot be processed by a traditional Relational Database Management System (RDBMS). It is generally used for big data and real-time web applications. Figure 3-15 and Figure 3-16 compare the performance and response speed, respectively, of the SM953 and SATA SSD in the RocksDB environment, the NoSQL database applied to Facebook®. The SM953 shows up to five times higher performance and response speed. This suggests that the SM953 can provide numerous advantages in the NoSQL environment. TPS (Transaction per Second)
SM953
15,000
21,825
20,000
19,852
Performance (IOPS)
SATA SSD 25,000
10,000 5,000 0
※ Refer to the Appendix for test conditions and workload. [Figure 3-14] TPC-C Performance Comparison
SATA SSD
RocksDB Performance
SM953
500,000
11,241
11,028
373,557
374,752
389,398
147,008
65,418
100,000
68,999
200,000
68,889
300,000
138,526
400,000
0 Bulk_load_random_order
Bulk_load_sequential_order
Multi thread single thread
Read Perf
Write Perf
※ Refer to the Appendix for test conditions and workload. [Figure 3-15] RocksDB IOPS Performance Comparison
16
Latency(us)
IOs 10,000,000,000
SATA SSD
5,000
350,000
1,000
1,800
800
1,200
500
350
200
90
Latency(us)
Multi Thread Read & Single Thread Write
SM953
140
Bulk load keys in sequential order
60
SM953
40
1
5,000
350,000
1,000
1,800
800
1,200
500
350
200
90
140
60
1 40
1 25
100
16
100
7
10,000
10
10,000
4
1,000,000
1
1,000,000
SATA SSD
25
IOs 100,000,000
16
Bulk load keys in random order
SM953
7
SATA SSD
10
IOs 100,000,000
4
SM953 White Paper
100,000,000
IOs 100,000,000
SATA SSD
Random Read
SM953
1,000,000
1,000,000 10,000 10,000 100
100
SATA SSD
SM953
50
IOs 100,000,000
6
3,000,000
700,000
350,000
1,60,000
70,000
35,000
1,400,000
Latency(us)
16,000
8,000
4,000
1,800
900
450
200
50
Latency(us)
100
25
12
6
500,000
300,000
160,000
90,000
50,000
30,000
9,000
16,000
5,000
3,000
1,600
900
500
300
160
90
50
30
16
9
5
1
1
1
1
Random Write
1,000,000 10,000 100
700,000
300,000
1,40,000
70,000
35,000
16,000
4,000,000
Latency(us)
8,000
4,000
1,800
900
450
200
100
25
12
1
1
※ Refer to the Appendix for test conditions and workload. [Figure 3-16] RocksDB Response Time Comparison
17
Conclusion
The SM953 is Samsung’s second NVMe product after the high-end level XS1715. It is a popular datacenter product that can provide consistently high performance at lower cost than the existing PCIe SSD. Since it uses the NVMe protocol, which is standardized to utilize the non-volatile memory performance such as NAND flash, unlike the traditional storage device (disk), it can provide three times higher performance compared to the existing SATA/AHCI product with the same NAND flash. Datacenter performance improvement using the SM953 varies according to the frequency of use of the storage by the host. If a multi-core, high-frequency and high-performance CPU is used, as well as the PCIe 3.0 high-performance interface, the best performance improvement can be achieved. Using the SM953, especially for applications such as server virtualization and database server, as described in Chapter 3, will deliver satisfactory system performance improvement. In addition, by combining Samsung’s genuine controller technology and NAND flash management technology, the SM953 can guarantee 0.9 DWPD, backed by a five-year warranty. With 6 watt of active power, the SM953 is the best product to lower a datacenter's TCO.
18
SM953 White Paper
Appendix
Specification PM853T Form Factor Capacity
Physical Dimensions
Host Interface
Performance* Power Consumption* Temperature
M.2
2.5"
480 GB
480, 960 GB
(22 ± 0.15) x (110 ± 0.15)mm (0.87 ± 0.01) x (4.33 ± 0.01)in., Top 5mm (0.2in.) Max, Bottom 1.5mm (0.06in.) Max
(100.2 ± 0.25) x (69.85 ± 0.25) x (6.8 ± 0.2)mm (3.95 ± 0.01) x (2.75 ± 0.01) x (0.27 ± 0.01)in. Label thickness is included in the thickness dimensions.
- Fully supported 1.0e, NVMe 1.1a compatible - 1 Namespace, 8 Queues - Support Atomic Write - PCIe Gen2/Gen3 x2/x4, INTx/MSI/MSI-X Sequential R/W - 1,750/850MB/s (8KB map) Random R/W - 260K/14K IOPS (8KB map) Active Read/Write: 5.5Watt /6.4Watt, Idle : 1.9Watt Operating : 0°C to 70°C (32°F to 158°F) Non-operating : -40°C to 85°C (104°F to 185°F) 5% to 95%, non-condensing 7~500Hz, 2.17Grms, 15min/axis (X, Y, Z) 1,500 G, duration 0.5m sec, Half Sine Wave 2.0 million hours
Sequential R/W - 2,200/1,400MB/s (8KB map) Random R/W - 300K/16K IOPS (8KB map) Active Read/Write: 4.7Watt /6.5Watt, Idle : 2.1Watt
Humidity Vibration Shock MTBF 480GB : 4,245/750TB, TBW 4,245/750TB (8KB Map) (Best/Worst) 960GB : 8,490/1,500TB (8KB Map) Weight 15g 63g * Actual performance may vary depending on use conditions and environment
Evaluating the Performance A. Considerations i. The product shall be sufficiently pre-conditioned for sustained performance. ii. Proper systems, threads and queues shall be set up to realize maximum performance by the product. (E5/i7-four-core or more, Thread x Queue ≥ 256)
19
B. Performance Evaluation Environment i. System: Dell™ PowerEdge™ R720 Server (RAID/TPC-C Evaluation), Dell PowerEdge R720xd Server (Jetstress Evaluation), HP® Z230 Workstation (Basic/Server Application Evaluation), Intel™ SR2612URR Server (RocksDB Evaluation), Dell T620 Server (Virtual Machine Evaluation) ii. Processor: Intel Xeon™ CPU E5-2690 @2.90GHz, Intel Xeon CPU E5-2670 @2.60GHz, Intel Xeon CPU E3-1230 v3 @3.40GHz, Intel Xeon CPU E5620 @2.40GHz, Intel Xeon CPU E5-2650 v2 @2.60GHz iii. OS: Windows® Server® 2012 R2/RHEL 6.5 Kernel 3.0 or higher (Inbox NVMe Driver) iv. Test Target: Physical Device, Full Range v. Test Tool: IOMeter 2006.07.27 (Windows), FIO 2.1.3 (Linux®), Benchmark Tool vi. Test Time: 1 minute per item
C. Performance Evaluation Items Performance Test Script Precondition-Sequential
128 KB Sequential Write (Density x 2)
Sequential Read Performance Test
Worker 1-256, Queue Depth by worker: 1-256 (Block Size=64KB/128KB)
Sequential Write Performance Test
Worker 1-256, Queue Depth by worker: 1-256 (Block Size=64KB/128KB)
Precondition-Random
Worker 1-256, Queue Depth by worker: 1-256 (Block Size=4KB/8KB)
Random Read Performance Test
Worker 1-256, Queue Depth by worker: 1-256 (Block Size=4KB/8KB)
Mixed (70/30) Performance Test
Worker 1-256, Queue Depth by worker: 1-256 (Block Size=4KB/8KB)
Random Write Performance Test
Worker 1-256, Queue Depth by worker: 1-256 (Block Size=4KB/8KB)
Server Application Workload Main Request Size
Sequential
Random
Read
Write
Web Server
4KB
25%
75%
95%
5%
Media Streaming
64KB
100%
0%
98%
2%
Exchange Email
4KB
0%
100%
67%
33%
Database OLTP
8KB
0%
100%
70%
30%
Decision Support System
64KB
0%
100%
100%
0%
Server Application Workload Web Server Application Server Database Server
Firmware Update Cautions: Firmware downloads destroy data. Therefore, before starting a firmware download, all data in the SSD should be backed up. Do not remove any SSD or execute new firmware download while an existing firmware download is in progress.
20
SM953 White Paper
1) Run “Windows Samsung NVMe Re-Drive.” 2) Click the “Firmware Tab” and then select the drive where the firmware download will be executed from the “Drives” list. 3) Select the F/W slot.
4) Click “Download” and then select the execution method. (In the following example, “Replace firmware image in slot 3 activate after next reset” has been selected.)
5) Specify the F/W path, open the firmware binary and then download the firmware.
Analysis Tool To analyze the PCIe SSD, use Teledyne LeCroy™ PETracer™ 7.0. For more information, visit the Teledyne LeCroy website at http:// teledynelecroy.com.
21
www.samsung.com/ssd