How to perform POC in depth with full flash distribution?

Full flash storage has become the upstart of current storage , From the perspective of the evolution of storage architecture , It is a general trend for software defined distributed storage system to replace disk array with traditional controller architecture . The full flash storage system with distributed architecture can give full play to the performance of flash media , And build an elastic infrastructure with cloud attributes , It is the development direction of all flash storage system in the future .

FASS yes TaoCloud Self developed full flash distributed block storage system , Since the official release of the product, a large number of scenarios have been carried out POC test , Here will be a description of the typical scene POC Multi angle comparative analysis of the test , Verify with data FASS Excellent performance in different hardware configuration environments .

Typical scenario POC Measured comparison

In actual measurement FASS More theoretical performance ：

• Play the role of theoretical performance in the actual test of the company 82%;
• In an internet test , Because there are fewer volumes , Not to optimal performance , put out a surge of energy 55% Theoretical performance ;
• In a certain AI Manufacturer testing , some Ceph product （ be based on RBD） Accessible 32 ten thousand IOPS,FASS In less 4 block SSD In this case , achieve 132 ten thousand IOPS,IOPS Difference between 4 Twice as many .FASS The best measured performance can reach the theoretical 80% above ,Ceph Give full play to the hardware performance 14% about .

FASS And Ceph Performance comparison

Ceph yes SDS More mature solutions in the industry , But limited by architecture design and other reasons ,Ceph It has obvious performance bottleneck in full flash environment .

The picture below is 3 Node FASS And optimized Ceph Contrast test ：

• It can be seen that IOPS Yes 5 times ～9 Times the difference ;
• The delay is more 18~73 Times the difference ;
• FASS Better performance than Ceph, Have higher IOPS, Lower latency .

FASS Ultra low delay ：

FASS Adopt parallel pipeline design , Will read and write I/O Evenly distributed in multiple nodes of each node CPU The core performs efficient parallel cooperative processing , Thus, the very low delay response of microsecond level in distributed system is realized .
Here's the picture , In different hardware scenarios , Excellent performance .

• stay 25G And above ,nvme Disk environment , Delay in 200μs within ;
• On the 10 Gigabit Network SATA Plate , Only 542μs Delay of .

FASS Enrich the agreement

stay FASS In different scenes , Support multiple protocols , Different protocols have different performance .

• FASS Support standards iSCSI Block storage protocol and high performance iSER visit , use RoCE Protocol to implement iSER visit ;
• FASS Currently adopted RDMA Transport（Infiniband or RoCE v2） Realization NVMeoF Agreement to access .
  Here's the picture ,100Gb/nvme/NVMeoF scene , It can reach tens of millions IOPS, At the same time, it has ultra-low delay, which is only 166μs.

FASS characteristic

Single volume performance

• stay FASS In different hardware environments , Single volume performance can play an ultra-high performance .
• With the improvement of hardware environment , The ultimate performance of a single roll also increases linearly ;
• stay 100Gb In the network environment , A single volume can reach 150 ten thousand IOPS.

Thin volume performance

• FASS Thin volume has the same performance as thick volume , Thin volume performance loss does not exceed 5%.
• Thin volumes can be expanded on demand , It is conducive to flexibly expand capacity without planning future capacity in advance , More cost savings .

QoS & Flow control

• FASS Support volume 、 Cluster level QoS Flow control , Flexible control IOPS Or bandwidth limit .
• colony QoS Can be set up ： Priority data repair （recovery Full speed ） Business or （recovery The speed limit ）

Recovery Performance climb
FASS recovery Fast performance climb , For business io Little impact .
As shown in the following example ,FASS Of Recovery The performance climb time can be controlled within 30 Within seconds .

There are many management tools
FASS Support CLI（ Command line management interface ）、REST API、Web GUI、FASS-CSI（ docking Kubernetes platform ） Configure the resources of the entire storage cluster 、 Access control 、 Data protection 、 Performance monitoring and other aspects of management .

FASS Key points for optimal performance

Hardware selection

• CPU： frequency 2.60GHz,L3 cache , 24MB,16core

• The Internet ：25Gb And above

• disk ：SATA SSD or NVMe SSD

• Memory ：128G

The system configuration

• Turn on NUMA, Turn off hyper threading
• Turn on CPU Performance mode
• If used raid Card mode , Must bring BBU
• operating system ：CentOS 7.6 Release edition
• Nodes need to be configured hostname、/etc/hosts file 、 Mutual confidentiality
• close selinux And firewalls

Software related

• CPU coremask ： from cpu Reasonable allocation of audit numbers
• Number of pressure coils ：Client Number *6
• Number of networks ： Management network 1, Front end network 2, Back end network *2
• fictitious Vip Number ：server Number = fictitious vip Number

POC Test request

“ This article is mainly to throw a brick to attract jade , adopt POC Comparison of measured data , And Ceph Product comparison and other aspects , It shows FASS High flexibility 、 Rich expansibility , Ultra high performance 、 Very low latency ,FASS In terms of performance 、 Cost performance and other advantages .”