Data Center Tech Blog

The rise and abundance of applications and devices at the edge of the Cloud has led to the massive growth of information. This pinpoints the incredible need for better and faster processing methods to profit from the captured data. As a result, there has been a growth of in-memory databases, analytics, caching and other applications.

Patric Chang

Senior Director at SanDisk’s Global Technology Ecosystem

Deploying In-Memory Processing

Even though in-memory processing has gotten popular, there are two limitations to be aware of when deploying:

Capacity limitations

  • Memory module capacity: Currently 64GB DIMMs are common, while 128GB DIMMs are expensive alternatives with very limited supply.
  • Server total memory capacity: There is a cap on total capacity as expressed by the maximal number of DIMM slots per server, multiplied by the maximal DIMM module capacity.

 

Cost limitations

  • Currently, commodity 64GB DIMMs are less than $1000 per unit. On the other hand, 128GB DIMMs are still uncommon and can be costly (e.g. more than $4,000 per unit).
  • For distributed applications, it’s possible to increase total memory with additional server nodes. The added expense of systems, processors, and networking; however, might make it cost prohibitive.

Western Digital has partnered and worked with Supermicro and ScaleMP to address this issue.

Using HGST high performance Ultrastar® SN260/SN200 PCIe NVMe™ flash drives, Supermicro® servers and ScaleMP software-defined memory (SDM-S) you can create a memory extension solution.

This solution will enable customers to deploy large system memory footprints using a combination of DRAM and HGST NVMe drives. It’s also transparent, as the operating system sees the solution as standard system memory. It’s almost as if much more DRAM was installed into the system.

Customers can now leverage better price/performance. Both for in-memory applications such as memcached, Apache Spark™, Redis®, Aerospike® and Apache Ignite™, as well as in-memory databases such as SAP HANA®, Oracle Database 12c, MySQL™, MongoDB®, ProgressDB. Customers can also run in-memory workloads in sizes that might otherwise be infeasible on commodity x86 hardware.

Performance Characterization

Initial performance tests for the memory extension solution used an in-memory caching software—memcached—on only 96GB of DRAM, expanded to a total of 768GB of memory using NVMe SSDs. In this configuration, it was able to deliver 98% of the performance (TPS) of a standard configuration with 768GB of DRAM as illustrated in figure 1.[i]

 

Figure 1: Memory Extension memcached performance characterization tests

Figure 1: Memory Extension memcached performance characterization tests

In a separate performance characterization test using an Online Transaction Processing (OLTP) workload (TPC-C) with MySQL for a multi-tenant use case, the memory extension solution using 96GB of DRAM to deliver 768GB of system memory, was able to deliver 80% and 74% of the performance (TPC-C) on two industry standard servers (ES-2637 v4 3.5GHz processor with 16 CPU threads) and (ES-2699 v4 2.2GHz processor with 88 CPU threads) respectively.

Figure 2: Memory Extension OLTP performance using MySQL TPC-C – 7 Instances x 66GB memory

Figure 2: Memory Extension OLTP performance using MySQL TPC-C – 7 Instances x 66GB memory

Figure 2: Memory Extension OLTP performance using MySQL TPC-C – 7 Instances x 66GB memory

We shall be running other pertinent workloads for our performance characterizations moving forward. From these two workloads, our tests show performance can be achieved by extending DRAM with NVMe flash to get as high as74% to 98% of DRAM performance.

Cost and Capacity Benefits

To illustrate the cost advantage, let’s take a 2-socket industry standard server with 768GB of DRAM. The left-most column in figure 3 shows the costs of a bare bone server, the CPUs and DRAM at $1,000, $2,400 and $6,912 respectively for a 768GB configuration.[i]

By using the Memory Extension Solution, we can use 96GB of DRAM and extend the memory to 768GB using an 800GB NVMe drive which resulted in a 30% reduction of the acquisition cost (shown in middle column).

In addition, by using a larger drive, we can extend the working memory to 1,536GB resulting in a 2X increase in memory for the same acquisition cost (shown in right-most column). The below example also illustrates the option for extending working memory by 8X using the memory extension solution.

The result from the memory extension solution?

You can reduce the required data center server footprint needed for large memory applications to 1/8. You also benefit from the corresponding cost savings from the reduction of server count, cooling and power consumption.

 

Figure 3: Cost Benefits of Memory Extension Solution

Figure 3: Cost Benefits of Memory Extension Solution

Memory Extension Benefits

These initial performance characterizations and budgetary estimates point to a very compelling proposition for customers looking into large memory configurations for in-memory processing requirements.

Three benefits of this solution are as follow.

First, with the memory extension solution, customers will be able to derive the full benefits of in-memory processing by increasing the working memory available.

Second, as the in-memory processing requirements grow customers can extend the working memory using high performance NVMe flash when the memory requirement grows above one or two TBs. No need to add additional servers. Customers now have more flexibility to scale based on CPU-utilization rather than maxing out on the amount of memory they can install in a server.

Third, by reducing the number of servers needed to support large memory configurations, customers are able to reduce the TCO and operational costs for in-memory deployments. In addition, customers have the added capability to scale their storage capacity.

Memory Extension Solution Showcase at CEBIT and World Hosting Day in Germany

Supermicro will be showcasing the Memory Extension solution with Western Digital and ScaleMP at CEBIT in Hannover, Germany from March 20-24 and World Hosting Day in Rust, Germany from March 27-31. If you are at these events, come by and see a demo at the Supermicro booth. We would love to show you how the different configurations may help you realize the full benefits of in-memory processing as described above.

 

[i] All results based on testing performed by ScaleMP in collaboration with Western Digital and SupermicroResults and performance may vary according to configurations and systems, including drive capacity, system architecture, applications, and test methodology.

[i] Costs are approximations based on available pricing information and may not be typical.

Key Considerations for Geo-Spreading with Object Storage

What in the World are WHQL, AQs and PCS?

subscribe blog