Fusionio Performance Test and Bottleneck Analysis

Source: http://www.hellodb.net/2011/06/fusionio-performance.html

Test environment: Dell r510, 2 x e5620, 24g, fusion-io iodrive 320g MLC

Test Tool: RedHat Linux 5.3, Oracle Orion 11

Test 1: 8 K random read with an iops of more than 5 W and a throughput of more than 400 m. When the response time reaches 4 W, the inflection point (more than 1 ms) appears and rapidly increases to the maximum value of 9.5 Ms.

######################################## ################

Test 2: 4 K random read, iops exceeding 7 W, throughput reaching 280 m, and response time keeping below 1 ms when iops reaches 7 W. As the concurrency pressure increases, iops fluctuates, the response time increases gradually.

######################################## ################

Test 3: 5000 k continuous reading, when the iops rapidly increases to 700, it reaches the throughput bottleneck (M). At this time, the response time is less than 10 ms, and then the concurrency pressure increases, there is no change in iops and throughput indicators, and the response time increases rapidly.

######################################## ################

Test 4: 8 K random write. The iops reaches 5.8, and the response time is 1 ms. As the stress increases, the iops gradually decreases, and the response time increases rapidly.

######################################## ################

Test 5: Hybrid read/write mode, 8 K random Io, 20% write, iops reaches 4.5 W, throughput is 360 MB, and response time is kept at around 1 ms when iops reaches. As the pressure increases, the response time increases rapidly.

######################################## ################

Test 6: Mixed read/write mode, 8 K random Io, 20% k continuous Io, 580 write, 5 W iops, MB throughput, and 1 ms response time when iops reaches 4 W, as the stress increases, the response time increases rapidly.

Performance analysis:

In this performance test, the fusionio card is read and written directly without any impact on the cache. From the test data, we can see that the random read performance of iodrive is very good and the 4 K Io performance is the best, it can reach 7 W iops, 8 K Io can reach 5 W iops, and the response time is stable below 1 ms. Random write performance is also very good, 8 K random write can reach more than 5 W. Iops, the response time is kept below 1 ms. Simulate 8 K random I/O and k continuous I/O of the database, and the iops can reach about 4 W. In short, iodrive performs very well.

Bottleneck Analysis:

There are two storage bottlenecks: iops and throughput. For traditional disks, The iops of a single disk is 150. If each Io is 1 MB at the maximum, we can calculate the Disk Throughput bottleneck of 150 MB (measured throughput is roughly 170-250 MB ). We can see that iops is the real bottleneck of the disk, random I/O is fatal to the disk, and throughput is usually not the bottleneck for the disk, so the disk is more suitable for the throughput system.

Compared with disks, fusionio can easily reach more than 5 W with a response time of less than 1 ms, while the throughput bottleneck is roughly between mb-700mb. The official data is not much different from the actual data. From the above data analysis, we can see that the iops of fusionio is very high, which usually does not become a bottleneck, and the throughput may become a bottleneck before iops. We can calculate the throughput of 128 MB. The IO size is 4800 K, and the iops is only 400. 8 K random reads. When the iops reaches 5 W, the throughput is close to MB. Although the throughput of a single fusionio card is larger than that of a disk, considering the price, fusionio is not suitable for throughput systems.

Observe the response time of fusionio. We find that the response time is stable at around 1 ms before the pressure reaches the bottleneck. When the performance bottleneck is reached quickly, the iops and throughput will not increase, at this time, a sudden change occurs in the response time, and then increases rapidly until it is unacceptable. Therefore, in actual use, the pressure must be controlled under the performance bottleneck.

Fusionio has high iops performance, so it is more suitable for random I/O read/write. In actual use, the I/O size can be appropriately reduced to convert the throughput bottleneck to the iops bottleneck, which is exactly the opposite of the disk system, the disk system should try to convert the iops bottleneck to the throughput bottleneck, such as increasing the I/O size, which is a very interesting topic.

With the emergence of flash storage technology, it will subvert the entire storage industry, and even change the design of storage systems such as databases in the future, which also gives us a huge space to imagine.

-EOF-

Official fusionio data: