2nd Generation Turbo Cloud

By Alasdair Lumsden on 8 Sep 2015


We launched our Turbo Cloud at the beginning of 2014, bringing online a fleet of brand new Dell R620 1u servers loaded with super fast hardware:

1st Generation Turbo Hardware
CPUs: 2 x Intel E5-2690 (14376 CPUMark Score)
Networking: 2 x 10Gbps
Disks: 8 x 480GB Samsung 840 Pro SSDs

These servers have served us well over the past 18 months, delivering blazing performance, especially when coupled with the SmartOS Hypervisor (The ZFS filesystem certainly knows how to use all that RAM for caching).

We’re now excited to announce the general availability of our 2nd generation Turbo cloud platform! The biggest change is that we’ve switched away from Dell to Supermicro, giving us more fine grained control over the internal components. We’re now also using a 2u chassis, allowing for up to 24 2.5" drives, significantly more than the 8 available in an R620:

2nd Generation Turbo Hardware
CPUs: 2 x Intel E5-2680v3 (16356 CPUMark Score)
Networking: 2 x 10Gbps
Disks: 10 x 800GB Intel S3610 SSDs, 2 x HGST SSD800MH.B 100GB SSDs (Write Cache)

Here you can see we’re now using Intel’s 3rd generation E5 CPUs, in addition to more DDR4 RAM. We’ve also more than doubled the amount of storage, and are using much faster Intel S3610 datacenter SSDs. But most interestingly, we’ve introduced 2 extremely high performance HGST 100GB write cache SSDs.

Addition of 2 Write Cache SSDs

To understand why we’ve added these, it’s useful to know that on a system there are two kinds of write operations – asynchronous, which can be written to disk whenever, and also synchronous writes, which must be written to disk now. With traditional spinning disks and a typical combined read/write workload, synchronous writes can take a long time to hit disk, degrading performance, especially for operations like database writes.

SSDs are meant to significantly help with the situation, as their read/write performance is substantially higher. However in reality, the latency of read/write operations starts to become highly variable under load, degrading performance. We’ve seen this on our 1st generation Turbo cloud – heavy synchronous write workloads can heavily interfere with heavy read workloads.

To solve this, in our 2nd generation Turbo cloud platform, we’ve introduced high performance 100GB HGST write SSDs. We allocate these within ZFS as "SLOG" (synchronous log) devices. ZFS will log synchronous writes to these SSDs now, and later flush them back to the main data pool whenever. The nice thing about whenever write operations is that they don’t interfere with read operations, as the scheduler can pick the best time to do them.

The end result is consistent, low latency high performance disk operations at all times. Result!

We hope you enjoy the new hardware and if you have any questions, please don’t hesitate to get in touch.

We’ll aim to post some benchmarks for these new systems over the coming months…