Next year, as part of a major server refresh, we should see productization of a new technology known as Intel® Optane™ DC Persistent Memory technology. This memory technology, supported on new Intel® Xeon® Scalable processors code-named “Cascade Lake,” represents a new class of memory and storage technology architected specifically for data center requirements. It offers a unique combination of high-capacity, affordability, and persistence. By so using this upcoming technology to move and maintain larger amounts of data closer to the processor, workloads and services can be optimized to reduce latencies and enhance performance.

My colleagues and I have been working on this technology with Intel since 2012; now our prototype labs are full of samples and it is pretty cool. Here are the key points that drew us initially to Intel® Optane™ DC Persistent Memory and which we believe are most relevant for VMware’s customers:


No exotic hardware connectors required: the technology will be available in memory modules that will plug into the DIMM slots of the new Intel based servers alongside standard DRAM.
Supports a standard load/store memory access model with byte-level granularity.
Provides persistence for new applications and virtual machines. (This is a very exciting development for changing the memory storage paradigm and VMware vSphere should be your choice for embracing these new applications.)
Provides a volatile mode for an easy uplift to existing workloads with no changes to software.
Details of Intel Optane DC PMEM

As we have discussed at VMworld, VMware is leading the charge to change the storage/memory paradigm with virtualization of byte-addressable persistent memory to new applications. That work will really payoff for future platforms with Intel Optane DC Persistent Memory. (We will re-post the details here later.) But, today, let’s focus instead on the newly announced volatile mode of this technology: “Memory mode.” When used in memory mode, the new Intel memory technology can greatly increase the memory capacity available to software in a platform when compared with the capacity of DRAM. This increase in capacity requires no changes to your existing software, operating systems, or virtual machines.

For example, consider a server with two processor sockets (“2-socket server”) with 24 DIMM slots running VMware vSphere. If all the DIMM slots are populated with 64-GiB DIMMs, then the maximum volatile memory capacity is 1536-GiB. If instead, those DIMM slots were populated with a mixture of 512-GiB memory modules of Intel Optane DC Persistent Memory and 64-GiB DRAM DIMMs using a future version of vSphere, the maximum volatile memory capacity could be 6144-GiB — without needing to make any changes to your virtual machines.

Of course, your mileage may vary and there is not a free lunch: although the latency of the Intel memory technology is good, nothing today can beat the latency of DRAM. But Intel makes up the difference in many cases by using DRAM in the other memory DIMM slots as a cache for the most frequently-accessed data, while the Intel Optane DC Persistent Memory provides large memory capacity. The caching is handled transparently by the Intel Xeon Scalable processor’s memory controller.

The net effect is that this is a very viable technology for memory-capacity-bound workloads. It can consolidate a larger number of virtual machines on a server than possible in the past without any changes to the virtual machines. It also allows running virtual machines with much larger DRAM requirements. For VMware’s customers, this means flexibility. You can dynamically repurpose server hardware for any workload, including performance-intensive machine learning and analytics applications. Not having to dedicate specially-configured servers for certain workloads allows you to maximize your infrastructure investments, doing more with fewer servers.


Extending Virtualized SAS Analytics Capacity

Rather than taking just our word for the benefit to VMware’s customers, VMware and Intel partnered with SAS, the global leader in enterprise analytics, to demonstrate the value. SAS provides a unified, open analytics platform replete with cutting-edge algorithms and AI capabilities.

Specifically, we performed tests using SAS® Viya®, which is a cloud enabled, in-memory, analytics engine that provides quick, accurate and reliable analytical insights. The scalability of SAS Viya made it the right choice to measure the number of models that could be simultaneously executed on each of two systems configured below. Please note that the “Cascade Lake” test system is configured with twelve 512-GiB Intel Optane DC Persistent Memory DIMMs for a total volatile capacity of 6-TiB.

Server Configuration for Test Measurements

A future version of VMware vSphere was installed on each host. Identical VMware virtual machines (VM) for the SAS® Viya® 3.4 workload were created with the below configuration. Each VM consumes 400-GiB of raw data for the analytics task. Most of the remaining memory on each VM is consumed by the application and related meta data. The VMs and models were run concurrently in batch for each test scenario (i.e., varying number of simultaneous modeling jobs) on both servers.

Workload Configuration for Test Measurements


The Results: Unlocking the Power of AI and Analytics

Today’s 2-socket server configured with 1.5-TiB of memory can normally run only up to 3 virtual machines concurrently (1 VM per disk/datastore) because more VMs would require more memory than is practical to populate. On a server using Intel’s “Cascade Lake” processors with 6-TiB of Intel Optane DC Persistent Memory in memory mode, our testing shows that we can run the same 3 VMs, or double, or triple the number of VMs – up to nine virtual machines (3 VMs per disk/datastore) while meeting an analyst’s turnaround expectations.

Result Measurements

Our measurements shown in the chart above means that more predictive models can run concurrently on tomorrow’s 2-socket server with virtualization than today’s. Let’s use an example to show why this can lead to better analysis. Today, an analyst focused on the probability of an individual contracting Type 2 Diabetes in a given population would have to create a one-size-fits-all predictive model because of the memory limitations of the current generation servers. This would be adequate, on average, but the analyst would have some difficulty properly predicting the tail-ends of this kind of distribution.

But, in the future, using the next generation Intel Xeon Processor and its new memory technology, the analyst can do finer-grain modeling of the outcome probability with the same class of server platform by creating a separate model for each combination of age group, gender, socioeconomic status, and geographic location, etc. And SAS® Viya® makes it simpler to automatically create an independent model for each target segment and virtualize the modeling job, thus increasing accuracy and relevancy in the analysis.

Modeling more variables can provide more accurate predictions for better decision support. Healthcare providers can use this information for preventative purposes for at risk individuals or to know who will develop diabetes and why? This can make a significant difference for both the individual’s health as well as the economy.


Closing Thoughts

The results show that a future version of VMware vSphere using the memory mode of Intel Optane DC Persistent Memory can triple the number of virtualized SAS analytics workloads that can run on a standard 2-socket server. Previously, you would need a 4-socket server to run these many models concurrently. Note again that no changes were made to the VM to extract this benefit. Your workloads may see the same benefit. We think it is worth investigating further.

You won’t have to wait long for these processor and memory capabilities from Intel in 2019. When Intel launches Intel Optane DC Persistent Memory, VMware intends support at launch of both memory and persistent modes.



Many experts in the VMware vSphere engineering team provided the foundation for support of this new technology aided by our fantastic prototypes team and qualified by our talented Hardware-Enablement and Quality Engineering teams. Generating the results and helping to write this blog required help from many external folks also:


SAS Michael Ingraham, Jeff Owens, Oskar Eriksson, Mitzi Krellwitz, Karina Estrada
Intel Suleyman Sair, Mike Ferron-Jones, Karthik Narayanan, Arakere Ramesh, Kristie Mann, Jack Vargas, Michael Strassmaier, Yashmeet Khopkar, Omid Meshkin, Victor Lee, Fal Diabate.



Memory Moment Video:
Intel Technical Details:
SAS Viya overview:
Type 2 Diabetes is a gradual affliction that surpises many when it starts having a noticeable effect on their health. So, some friendly advice from a Type 1 Diabetic of 40 years: get tested now. See