Optane essentials, Part 2: Memory modes and optimal workloads

This article is part of the Technology Insight series, made possible by funding from Intel.
________________________________________________________________________________________

In Part 1, we covered the media foundations of Intel Optane technology and the different implementations of that media into three products: Optane Memory, Optane SSDs, and Optane DC persistent memory modules (DCPMMs). We also touched on how Intel software is a key ingredient in turning 3D XPoint into Intel Optane. We’ll round out our 101 with a brief look at configurations and modes.

Why? While it’s true that Optane DCPMMs require 2^nd Generation Xeon Scalable processors or later, much of the Optane magic hinges on how software – especially in settings such as virtualized cloud, AI/analytics, and HPC –  can make optimal use of Optane when that media is configured into given modes.

Optane Modes

DCPMMs can be configured into three possible modes: Memory Mode, App Direct Mode, and Dual Mode. Note that DCPMMs do not replace DRAM; you still need some DRAM in the system. But how much DCPMM capacity you complement the DRAM with will depend on your mode and specific application needs.

Memory Mode turns DRAM into an L4 cache. The cache is non-addressable and doesn’t show up in system memory counts, so all user-addressable capacity is the sum of the DCPMM capacities. No programming is needed for Memory Mode, but data contained in DCPMMs is volatile, just like with DRAM. Memory Mode provides a large memory pool with low but effective DRAM cache investment.

App Direct Mode allows the system to handle DRAM and DCPMM resources independently, so operations that require top speed can address DRAM and the rest can rely on larger Optane resources. Under App Direct, data in DCPMM remains persistent, which can be very helpful in minimizing large configuration reload times following a power cycle or reset. However, applications must be optimized to take advantage of App Direct. Some are already, and more are coming. Also, some user programming may be needed.

Dual Mode is a sub-set of App Direct Mode that allows some DCPMM resources to operate in Direct Mode and the remainder in App Direct.

DCPMM modes, especially the first two, tend to get the most attention for performance reasons. However, if we turn our attention to Optane SSDs, we can employ Intel Memory Drive Technology back across the CPU’s PCI Express bus. In effect, Memory Drive is much like Memory Mode, including in having data volatility, only Optane SSDs are serving in place of DCPMMs. Naturally, there’s more latency involved in reaching out to SSD resources, but the size of memory pool possible can be relatively gargantuan. As a way to contain costs on projects that require memory capacity above all else, Memory Drive can be a life-saver.

Which workloads are best for Optane?

Speaking generally, Optane media excels when working with large volumes of real-time data that require low-latency access. Taking that one step further, we might add the following characteristics:

• Datasets requiring mass amounts of memory space, especially on the order of terabytes — more than might be physically or economically feasible with DRAM alone.
• Workloads that are memory-bound within virtual machine limits. IT managers often divvy up available hardware resources among different application owners, and overprovisioning of memory for these applications is a very frequent issue. Optane DCPMMs can provide enough memory to satisfy all VM owners without having to add additional servers. Alternatively, DCPMM capacity can allow IT to reduce its number of physical servers running virtualized environments.
• Applications, often mission-critical, that benefit from data persistence. When a system suffers an outage or even a planned reset, reloading data into memory can take an inordinately long time, potentially costing users/owners money and violating SLA requirements. Data persistence, as in App Direct Mode, can dramatically reduce reboot/reload times.
• Systems and applications with mass amounts of data writing per day. Optane’s markedly higher endurance makes it a more cost-efficient fit in environments where storage demands are likely to burn through conventional NAND-based drive options.

As mentioned previously, DCPMMs can be configured for either persistent or volatile operation and features latencies that are near DRAM in performance. This lends DCPMMs to deployment with workloads such as:

• SAP HANA
• Real-time analytics
• Database and/or database cache
• All virtualized and hyperconverged infrastructure memory extension
• Live streaming and hot video-on-demand workloads
• Apache Spark
• Major software packages, both open source and not, for data centers (may require optimization for Optane)

Returning to Optane SSDs, especially when running in a Memory Drive configuration, the following workload types perform particularly well:

• HPC Flex memory
• KVM memory extension
• Spark, k-means clustering
• Cloud service provider (CSP) use cases

For non-volatile workloads, you might consider Optane DC SSDs in applications such as:

• Ceph block/object
• VMware vSAN
• Hadoop YARN temp
• Cisco HyperFlex
• Microsoft SQL

Essentially, any storage application that suffers from I/O bottlenecking is a solid Optane SSD candidate. Also, if server monitoring tools reveal that CPU utilizations are low, this may be a sign that applications and/or VMs are memory-bound and could benefit from DCPMMs.

Wrapping up

And that’s about it for the Optane basics. We expect the technology to slowly make more inroads in the client space, but today it remains a clear data center play. Exactly how much cost advantages and performance gains can be provided by Optane implementations will depend on an organization’s specific needs and workloads.

But the ability to bring storage so much closer to the CPU, to the point that it becomes indistinguishable from memory, is compelling and certain to open new opportunities for application developers and enterprises alike.