Part Two – A New Approach | Hardware Matters in HCI
In Part 1 of this article we discussed the world of HCI today, or HCI 1.0. We left off with this question in mind: Is there a server out there that can meet what hyper-converged infrastructure should be, and not just what it has been in HCI 1.0?
With our vast history in performance-based SANs, we asked ourselves what if we squeezed the performance of an enterprise storage system into a server? The result – A new approach to server infrastructure where compute AND storage needs have been considered is here – FabricXpress™ (FX) is delivering enterprise scale with hyper-converged infrastructure simplicity.
How does FabricXpress deliver on this value to the Enterprise market? The greatest driver is that FabricXpress has the ability to support up to 72 NVMe SSDs per server and a unique midplane design that allows for a larger PCIe lane count to these drives allowing for the throughput and low latency required by enterprise applications.
Clearly, any HCI offering should leverage NVMe to its maximum. So, we return to the question – why design, build and offer a server architecture with that many NVMe SSDs if only a few can easily accommodate the I/O workload in one server? The answer is simple: Performance and Price. This is what Axellio brings to the HCI world – HCI 2.0.
Performance – NVMe
It is well known that NVMe SSDs are very fast storage devices, especially compared to the mainstream SSDs of just 5 years ago, never mind mainstream hard disk drives (HDDs). A single high-end NVMe SSD delivers up to 800K IOPS and up to 3.3 GB/s of I/O throughput. These numbers would have been considered astronomical for a single SSD just 5 years ago.
Upon closer examination, however, you’ll find that these are read-only metrics; once you look at the write metrics, the performance – especially IOPS – falls significantly. Write performance is delivered up to 160k IOPS for random small-block writes and 2-3 GB/s for sequential large-block writes. Still, these IOPS metrics are over 10x superior to an SSD of five years ago, and nearly 1000x superior to a traditional HDD of 5 years ago. NVMe SSDs are truly transformational in terms of block storage devices.
Undoubtedly, given the efficiency of the NVMe protocol and the never-ending optimization of NAND flash controllers in SSDs, the modern NVMe SSD is truly a performance ‘beast’ that is changing the landscape of how data can be accessed, especially in HCI.
Performance – Number of Devices Matters
Providing the capability to hold significantly more NVMe SSDs per server – 72, 96, 128, up to 144. This is three to six times the SSDs as one normally finds in the aforementioned popular servers. The ability to use many more NVMe devices in one server changes the game.
In the FX architecture, the I/O load is distributed across many more devices than other architectures, which translates to a lighter load per device. This means that each CPU core in the FX architecture is managing a lower queue depth per device, thus freeing up cycles for applications to consume. This not only makes them more efficient but also produces a lower, more predictable, response time (I/O latency), further enhancing application performance. In databases, for example, this translates to more transactions per unit time, with less latency per transaction.
Price – I/O Distribution
In particular, as the overall system load is distributed across more devices, so are the write I/Os. So each individual device may carry a lower endurance rating while still providing a long service life. It is interesting to note that many of the ‘scale-out’ databases use this technique, which is called ‘sharding’. It is the key technique to enable enterprise efficiency while using cost-effective components.
Essentially, by spreading the load over more devices, each device carries a smaller burden of the overall workload. Therefore, a more cost-effective device can be leveraged, staying within its native performance envelope, while still providing system level enterprise capability. But hang on, what about power fail handling to enable caching at the device level? Isn’t that important? Again the answer is simple – yes, it is.
To provide that capability, the FX architecture implements off-SSD power-fail handling by providing additional stored power (inside the server) and signaling that enables the SSD to gracefully take the time needed to flush data from its cache to NAND, to safeguard the data. This capability means that the SSD itself need not carry this feature, which further optimizes the cost.
Price – Smaller Clusters of Larger Servers
Lastly, smaller clusters of larger servers can ultimately provide a lower total cost of ownership (TCO). With the ability to support the I/O of storage-intensive workloads with a larger number of NVMe devices, that means you can load larger core count CPU into the system as well.
In a traditional HCI configuration comprised of a large cluster of small servers – you end up with sprawling IT and underutilized CPU cores that you are still paying for. To handle heavy storage workloads, you add multiple servers with small core counts and small amounts of storage devices. When you need more data storage you would “just add a shelf.” This can get out of hand and inefficient rather quickly.
With Axellio – up to 72 NVMe SSDs per node and 44 CPU cores per node – never pay for underutilized CPU and unnecessary licensing costs just to maintain the storage capacity you need.
Enterprise Scale | Hyper-Converged Simplicity
In summary, the FX architecture enables a tier-1, enterprise-grade, HCI workload-carrying capability at a significantly lower price. It provides the benefits of both lower price and higher performance, compared to traditional servers. It also provides a TCO improvement when considering power, cooling and rack space within an on-premises datacenter or co-lo – since deploying one FX often means the ability to recapture several rack U of space, which would otherwise be populated by ordinary servers.
The FX, with its superior architecture, is the optimal server on which to implement and deploy enterprise workloads in a simple and efficient hyper-converged infrastructure.