Ceph Performance Boost with NVMe SSDs
Many customers value the high level of performance of our servers. Along with powerful processors and fast networking, our distributed SSD storage cluster contributes substantially to superior speed. After thorough testing in our lab, we recently gave it an extra performance boost:
- Upgrading the Ceph storage platform
- Reinstallation of the storage nodes with Ubuntu 16.04 LTS
- Enhancement through super fast NVMe SSDs
Upgrading the Ceph storage platform
Ceph forms the basis of our distributed storage cluster. The open source solution ensures that your data is stored in a multi-redundant way and by that always available. Distributing read and write operations across a large number of storage nodes increases speed significantly. The physical separation of compute and storage nodes enables high-availability setups and allows storage capacity to be scaled virtually infinitely.
In order to ensure smooth operation, we rely on Ceph's "LTS" releases ("Long Term Stable" releases). Using a widely supported version increases reliability and minimizes the response time in case of an issue. After comprehensive testing in our lab we have updated the storage cluster from "Hammer" to the latest LTS release "Jewel". Impressive fact: Compared to Hammer, a performance increase of up to 35% has been measured using Jewel.
Reinstallation of the storage nodes with Ubuntu 16.04 LTS
Most of our servers are running Ubuntu, and here too we rely on versions with long-term support. While the productive systems were running smoothly with version 14.04 LTS, we have thoroughly tested the newer Ubuntu 16.04 LTS as well as the replacement process in our lab. During the last few days we have reinstalled all of our storage nodes with 16.04 and thus will continue to benefit from current software and long-term support.
Enhancement through super fast NVMe SSDs
When using Ceph, all data is stored in the so-called journal first. In a second step, data gets transferred to its actual storage location. In many cases SSDs are used for the journal to speed up write operations while data is then stored on magnetic hard disks.
At cloudscale.ch however, we use SSDs exclusively (except for bulk storage) to also ensure read access at lightning speed. To further optimize our setup, we have enhanced our storage nodes through NVMe SSDs. These serve as a super fast journal and speed up write access even more compared to the current SSDs. This new configuration also eliminates the "double-write penalty": Thanks to the separate NVMe journal data is written to the SSDs just once, which almost doubles data throughput.
After carefully planning every step, we were able to implement all these improvements in operation – probably without you even noticing. Of course, all customers automatically profit from the increased performance so there is no need to modify or replace your current servers.
At lightning speed,
your cloudscale.ch team