ユーザーズマニュアル目次Figures41.0 Revision History52.0 Supporting Documentation53.0 About This Guide54.0 Overview54.1 What Impacts SSD IO Performance64.2 Queue Depth and Latency74.3 Why Mixed Workload Is Important74.4 Drive Endurance84.5 Selection of RAID Controller85.0 RAID 195.1 Test System Specifications¹91. The system was selected to make sure the performance of the RAID card and the SSDs would not be inhibited by the server.92. Hyper-Threading is disabled in this test system specifically due to additional latency introduced during benchmark testing. In any practical application, Hyper-Threading would NOT be disabled.93. In the configuration of the RAID set, No Read Ahead and Write-through are used due to the speed of the SSDs. Read and write caching was designed for use with HDDs. Caching with SSDs introduces additional overhead thus interfering with the SSDs perf...94. One thread, or worker, per drive was used in order to simulate the manner in which many applications utilize storage and also to attempt to saturate the communication channels to the SSDs.95.2 Intel® SSD DC S3500 Series in RAID 1 Performance Characterization Data10Figure 1. RAID 1 Random 100% Write @ 4KB Transfer Size with Average Latency11Notes: Figure 1 - The write performance of the two drive RAID 1 set matches the write performance of a single Intel® DC S3500 drive. This indicates very low latency introduced by the RAID controller.11As more drives are added, the write performance scales linearly. At four drives, the performance is 2x that of a single drive, at six drives, it is 3x higher and at eight drives it is 4x. This is true at all queue depths tested and at transfer sizes ...11At a queue depth of 1, the average latency for 100% write at 4KB transfer size is less than 200 µs. Latency increases as the queue deepens, ending at 1.4ms for a queue of 8. It is interesting to note that latency is not affected by the number of drives.11Figure 2. RAID 1 Random 70% Read @ 4KB Transfer Size with Average Latency11Figure 3. RAID 1 Random 90% Read @ 4KB Transfer Size with Average Latency12Figure 4. RAID 1 Random 100% Read @ 4KB transfer size with Average Latency125.3 RAID 1 Consistency13Figure 5. RAID 1 Maximum Latency for 2-drive and 8-drive Configurations135.4 RAID 1 Performance Conclusions136.0 RAID 5146.1 Test System Specifications14The system used for RAID 5 testing was identical to the system used for RAID 1 testing except the following changes:14 2x Intel Xeon E5-2680 8-core CPUs (2.7 GHz)14 3x up to 8x Intel SSD DC S3500 Series 800GB drives14Note: For this test, 800GB drives were used. The rated performance of the Intel SSD DC S3500 Series drive in 800GB, 600GB, and 480GB capacities are nearly identical, per internal Intel testing.146.2 Intel SSD DC S3500 Series in RAID 5 Performance Characterization Data14Note: The scale of the IOPS charts is variable in order to clearly show the change as drives are added.14Figure 6. RAID 5 Random 100% Write @ 4KB Transfer Size with Average Latency15NOTES: There are gains in write performance as drives are added to the RAID 5 set. The change at queue depth 1 from three drives to six drives is approximately 58% increase in IOPS. For eight drives, the change is 97% increase in IOPS over the three ...15At a queue depth of 1, latency increases as more drives are added to the RAID set, most likely caused by the additional overhead of calculating parity and striping across more drives. This increases as the queue deepens.15Figure 7. RAID 5 Random 70% Read @ 4KB Transfer Size with Average Latency15Figure 8. RAID 5 Random 90% Read @ 4KB Transfer Size with Average Latency16Figure 9. RAID 5 Random 100% Read @ 4KB Transfer Size with Average Latency16Notes: Figures 7, 8, 9 - As the workloads become more read intensive, there is a steady increase in performance both as drives are added and as the queue deepens.16Figures 7, 8, 9 – As read percentage increases, the exponential increase in latency is not as prominent with deeper queues. This is due to the speed at which reads are performed.166.3 RAID 5 Consistency16Figure 10. RAID 5 Maximum Latency for 3-drive and 8-drive Configurations17/ /17Intel internal testing, October 2013176.4 RAID 5 Performance Conclusions177.0 Summary18To summarize:18 In both RAID 1 and RAID 5, the Intel SSD DC S3500 Series drive shows excellent scalability, performance, and consistency.18 Very little latency was introduced by the RAID controller in RAID 1. In RAID 5, the overhead and latency are slightly higher.18 In random, mixed read/write workloads, SSDs perform significantly (as much as 100 times) better than HDDs in a similar situation.18 With this RAID controller, there is the possibility of greater performance by adding more than eight drives in both RAID 1 and RAID 5 configurations.188.0 Appendix198.1 RAID Levels19サイズ： 624KBページ数： 19Language: Englishマニュアルを開く
ユーザーズマニュアル目次Figures41.0 Revision History52.0 Supporting Documentation53.0 About This Guide54.0 Overview54.1 What Impacts SSD IO Performance64.2 Queue Depth and Latency74.3 Why Mixed Workload Is Important74.4 Drive Endurance84.5 Selection of RAID Controller85.0 RAID 195.1 Test System Specifications¹91. The system was selected to make sure the performance of the RAID card and the SSDs would not be inhibited by the server.92. Hyper-Threading is disabled in this test system specifically due to additional latency introduced during benchmark testing. In any practical application, Hyper-Threading would NOT be disabled.93. In the configuration of the RAID set, No Read Ahead and Write-through are used due to the speed of the SSDs. Read and write caching was designed for use with HDDs. Caching with SSDs introduces additional overhead thus interfering with the SSDs perf...94. One thread, or worker, per drive was used in order to simulate the manner in which many applications utilize storage and also to attempt to saturate the communication channels to the SSDs.95.2 Intel® SSD DC S3500 Series in RAID 1 Performance Characterization Data10Figure 1. RAID 1 Random 100% Write @ 4KB Transfer Size with Average Latency11Notes: Figure 1 - The write performance of the two drive RAID 1 set matches the write performance of a single Intel® DC S3500 drive. This indicates very low latency introduced by the RAID controller.11As more drives are added, the write performance scales linearly. At four drives, the performance is 2x that of a single drive, at six drives, it is 3x higher and at eight drives it is 4x. This is true at all queue depths tested and at transfer sizes ...11At a queue depth of 1, the average latency for 100% write at 4KB transfer size is less than 200 µs. Latency increases as the queue deepens, ending at 1.4ms for a queue of 8. It is interesting to note that latency is not affected by the number of drives.11Figure 2. RAID 1 Random 70% Read @ 4KB Transfer Size with Average Latency11Figure 3. RAID 1 Random 90% Read @ 4KB Transfer Size with Average Latency12Figure 4. RAID 1 Random 100% Read @ 4KB transfer size with Average Latency125.3 RAID 1 Consistency13Figure 5. RAID 1 Maximum Latency for 2-drive and 8-drive Configurations135.4 RAID 1 Performance Conclusions136.0 RAID 5146.1 Test System Specifications14The system used for RAID 5 testing was identical to the system used for RAID 1 testing except the following changes:14 2x Intel Xeon E5-2680 8-core CPUs (2.7 GHz)14 3x up to 8x Intel SSD DC S3500 Series 800GB drives14Note: For this test, 800GB drives were used. The rated performance of the Intel SSD DC S3500 Series drive in 800GB, 600GB, and 480GB capacities are nearly identical, per internal Intel testing.146.2 Intel SSD DC S3500 Series in RAID 5 Performance Characterization Data14Note: The scale of the IOPS charts is variable in order to clearly show the change as drives are added.14Figure 6. RAID 5 Random 100% Write @ 4KB Transfer Size with Average Latency15NOTES: There are gains in write performance as drives are added to the RAID 5 set. The change at queue depth 1 from three drives to six drives is approximately 58% increase in IOPS. For eight drives, the change is 97% increase in IOPS over the three ...15At a queue depth of 1, latency increases as more drives are added to the RAID set, most likely caused by the additional overhead of calculating parity and striping across more drives. This increases as the queue deepens.15Figure 7. RAID 5 Random 70% Read @ 4KB Transfer Size with Average Latency15Figure 8. RAID 5 Random 90% Read @ 4KB Transfer Size with Average Latency16Figure 9. RAID 5 Random 100% Read @ 4KB Transfer Size with Average Latency16Notes: Figures 7, 8, 9 - As the workloads become more read intensive, there is a steady increase in performance both as drives are added and as the queue deepens.16Figures 7, 8, 9 – As read percentage increases, the exponential increase in latency is not as prominent with deeper queues. This is due to the speed at which reads are performed.166.3 RAID 5 Consistency16Figure 10. RAID 5 Maximum Latency for 3-drive and 8-drive Configurations17/ /17Intel internal testing, October 2013176.4 RAID 5 Performance Conclusions177.0 Summary18To summarize:18 In both RAID 1 and RAID 5, the Intel SSD DC S3500 Series drive shows excellent scalability, performance, and consistency.18 Very little latency was introduced by the RAID controller in RAID 1. In RAID 5, the overhead and latency are slightly higher.18 In random, mixed read/write workloads, SSDs perform significantly (as much as 100 times) better than HDDs in a similar situation.18 With this RAID controller, there is the possibility of greater performance by adding more than eight drives in both RAID 1 and RAID 5 configurations.188.0 Appendix198.1 RAID Levels19サイズ： 624KBページ数： 19Language: Englishマニュアルを開く
