The future SSD is here—PCIe/NVMe boosts Ceph performance

Ceph Day.jpg

Ceph shows a strong trend as opensource scale out storage adoption in worldwide market and we are observing strong customer requirements for high performance storage, from must-have SSD as Journal and caching to all SSD flash solutions, ranging from CSP, FSI, Telecom, HPC/government etc to OEM/ODM.

On the weekend of October 18, as the top-level Ceph community and industry conference on Ceph technology in china, themed “Ceph: The future of storage”, 2015 Shanghai Ceph day attracted 33 companies and over 140 developers, IT experts, academic leaders, business and technical managers etc, Intel delivered 1 opening and 4 key technical presentations along with Redhat, Suse, Mellanox, H3C and other industry partners delivered 10 other technical sessions.

On behalf of Intel, I along with two other Intel engineers presented “SSD/NVM technology Boosting Ceph performance”, see attached pdf, we propose first ever all SSD Ceph configuration, combination of 1x NVMe SSD (Intel P3700 800GB) as Journal + 4 x Low cost high capacity SATA SSD (Intel S3510 1.6TB) as OSD data drives, this configuration has dramatically increased random write performance ~100K iops (4 node cluster), which is >32x compares to SSD as journal + 40 HDDs configuration, or in order to reach 100K iops, you need total of 1300 HDDs, this has not counted power consumption, HDD fail rate, space, maintenance cost etc, you can image the total cost ownership (TCO) for using all SSD would be dramatically lower than HDDs…

We also presented Intel iCAS + Intel PCIe/NVMe SSD P3700 accelerates Ceph performance.

In addition, I propose three Ceph configurations: 1)standard/good Ceph configuration, PCIe/NVMe SSD as Journal and caching, plus HDDs as data drives, the ratio is 1:16/20, example is 1 x Intel P3700 800GB SSD + 20 HDDs, P3700 as both Journal and caching (with Intel iCAS), 2)advance/better configuration: NVMe/PCIe SSD as Journal + large capacity SATA SSDs, example is above 1xP3700 + 4 S3510; 3)the best performance configuration will be ALL NVMe/PCIe SSDs, example is 6xP3700 2TB SSD per node.

SSD vs HDD.png