Intel® Optane™ technology drastically reduces latency to data at the memory, device, and computing system level.
From the very beginning, computing systems have exploited the tendency of algorithms to access some data more than other data (temporal locality), and to access data stored close together (spatial locality), to increase performance and reduce cost. Data accessed more often is held in faster, smaller memories, while bulk data accessed less frequently is held in storage, costing much less per bit. This interplay between the memories within a computing system—and software running on that system—is commonly called the memory and storage hierarchy.
At every level in this hierarchy, both bandwidth and latency of data access matter. Over the years, sustained improvements in existing memory and storage technologies have produced continual increases in bandwidth, while latencies have remained much more constant. With the introduction of Intel® Optane™ technology as both memory and storage, this hierarchy is seeing its most significant addition since the introduction of NAND SSDs—or arguably since the introduction of DRAM itself. This paper explores the historically low latency oﬀered by Intel® Optane™ SSDs first, and then delves into the system level advantages of Intel® Optane™ DC persistent memory—the same underlying Intel® Optane™ media deployed as system memory.
Intel® Optane™ DC SSDs Significantly Reduce Storage Latency
Intel® Optane™ technology delivers unprecedented reductions in storage latency, reductions not just in the average latency an application sees but also (even more) in the infrequent longer latencies that occur under heavy system loads.
Figure 1 shows a comparison of two SSDs, a leading NAND-based Intel® SSD DCP4610 and an Intel® Optane™ DC SSD P4800X.1 This complex latency vs. load chart requires some explanation.
Read Latency Most Often Determines System Performance
The x-axis is the total mixed read/write IOPs delivered by the SSD under test. For each load (IOP level delivered) the latency of the read I/Os is measured using a ﬂexible I/O tool (FIO), to data return, and plotted using the y-axis. We measure read latencies because software often needs the data requested to make progress, that's why it reads the data in the first place. Writes, on the other hand, are quickly guaranteed persistent by most SSDs, which hold them in a SRAM, assured to be ﬂushed to non-volative memory, even in the event of a power failure. So it is read latency that most often determines system performance.
The performance for each SSD is plotted as a family of lines (solid blue for Intel® Optane™ SSD and dashed green for NAND SSD). The lines represent average latency measured and the longer, less frequently seen read latencies (often called tails since they appear that way in the right side of a histogram of latencies). The time to complete a read depends on many factors (the memory technology, how busy the SSDs are, etc.) and so varies. It's important to understand average latency and also the latency distribution. For each SSD, the bottom line is the average—or 50th percentile—latency.
The next line up is the 99th percentile latency, and so on up to the top line representing 99.999th percentile latency. 99.999th percentile latency corresponds to the next-to-the-slowest read out of 100,000 reads—an infrequent event that matters to some applications.
Finally, don't miss the labeling on the y-axis, a semi-log plot so each grid line represents a 10x diﬀerence from the previous. Now that Figure 1 is understood, we can use it to draw a few important conclusions.
Intel® Optane™ Technology Breaks Through the NAND Barrier
Intel® Optane™ technology is built on a completely new memory media that supports write in place, is byte addressable like DRAM, non-volatile like NAND, and has a read/write latency between the two. Intel® Optane™ technology combines Intel® Optane™ memory media with Intel® controllers, software, and system interconnects that can be deployed as memory or storage.
Intel® Optane™ Memory Media Provides Faster Access
Intel® Optane™ memory media delivers data significantly faster than NAND media. And the Intel® Optane™ SSD has been designed to provide that advantage to the rest of the computing system through a hardware-only SSD read/write path through the SSD controller, unlike the firmware-involved path found in NAND SSD controllers.
The average latency line for each SSD (the bottom lines in Figure 1) shows that the Intel® Optane™ SSD delivers greater than 10x latency advantage (measured using a ﬂexible I/O tool, or FIO) at every 70/30 read/write IOPs level shown. This means data is returned, on average in just 1/10th the time. Intel® Optane™ memory technology is byte addressable and allows rapid, in-place writes. NAND requires large block reads and a very large block erases before relatively slow writes. Data returned faster means less time waiting and faster application execution time.
Specification sheets for SSDs often quote nominal or idle latency. Figure 1 shows this latency as the left most point of the average latency line, this is the queue depth equals one (QD=1) point in the curve. When measuring this point, FIO asked for a single 4 KB read or write, and waited until that transfer completed before asking for another, using just a single thread. Notice for QD=1, the Intel® Optane™ SSD delivers far higher IOPs than the NAND SSD. This is precisely because it oﬀers lower latency. In fact, this is the impact of that lower latency to throughput. This also means that QD=1, or really any constant QD, is hardly a fair comparison point, since the Optane SSD is completing far more work (more IOPS) than the NAND SSD. SSD latencies should be compared under the same load like Figure 1.
Using this media advantage, Intel® Optane™ SSDs avoid long waits for writes to complete and the garbage collection inherent in NAND SSDs. This means that Intel® Optane™ SSDs don't face the background writes, and the delay to reads, seen by NAND SSDs. These advantages translate to significant quality of service (QoS) advantages described above. QoS is visible in the charts through the height and shape of the family of lines extending above the average latency line for each SSD. The NAND SSD has a 99.999th percentile latency that is more than 10x the average latency, while the Intel® Optane™ SSD latency distribution stays within that 10x threshold. More importantly the Intel® Optane™ SSD lines, at all percentiles, are much ﬂatter than the NAND SSD. This more ideal curve means that the Intel® Optane™ SSD delivers steadier performance, even under heavy loads.
For many applications, lower SSD latency, and better SSD QoS translates to better performance. For some applications, we see the lower average latency result in decreased runtimes, with less total time spent waiting for storage access to return. In this case, the Intel® Optane™ SSD delivers a better user experience. For other applications where user responsiveness requirements exist, and where each user visible operation requires multiple accesses, we see the improved QoS result in an ability to support many more users before the responsiveness requirements are exceeded. In this case the Intel® Optane™ SSD allows the system to support more users for a cost advantage.
Software Latency: The New Frontier
When accessing SSD resident data—even for modern Intel® Optane™ DC SSDs—there are several software layers in the operating system, between the drive and the application, that contribute to latency (see Figure 2). This software is commonly called the “storage stack.”
When an application initiates a read or write, that request is handed oﬀ to the operating system. Together, the instructions executed in the storage stack can take a total of up to 4–10 microseconds (μs), or even longer, of CPU work. This time can be as large as or larger than the time required to move data over the PCIe* input/output (I/O) bus using the NVMe* protocol, plus the time required by the SSD itself to read data from the storage media.
Time spent by the operating system isn't just time waiting, it's also time in which the CPU is busy. Imagine working on a system with a 3 GHz CPU. In this example, 1 microsecond is equivalent to 3,000 clocks. That means 10 microseconds of software latency equals 30,000 clocks spent not executing the user application. No matter how fast the hardware is, that software overhead is about the same. To get around that overhead, a new approach is needed.
Bypassing Software Overhead with Intel® Optane™ DC Persistent Memory
By deploying Intel® Optane™ technology as persistent memory instead of an SSD, much of that software overhead can be eliminated (see Figure 3). Intel® Optane™ DC persistent memory directly attaches to the memory channel instead of the PCIe bus. Intel® Optane™ DC persistent memory is a new tier of memory that sits on the faster memory bus, in a module form factor, but has some qualities of storage (persistence).
Intel® Optane™ DC persistent memory avoids software overhead by enabling mapping of persistent data for access directly from the CPU with memory load and store instructions—bypassing the operating system storage stack when reading and writing data. The operating system is directly involved in setting up access, providing a pointer to the location of the data, but need not act for regular reads and writes, resulting in much faster, direct load/store access to large capacity byte addressable space for working data.
The highly optimized architecture of Intel® Optane™ DC persistent memory module reduces latency considerably. This module contains a specially designed memory controller with a highly optimized hardware only read/write path and the ability to connect to the customer memory controller in the CPU itself.
With the new media and optimized design, idle average latency of less than 100 ns if the data is found in DRAM, and up to 340 ns if it is found in Intel® Optane™ DC persistent memory (in Application Direct Mode).1 The latency is specified as a range because these two memory types work together as a hierarchy (see Figure 4).
In addition, because persistent memory is treated by the system as memory rather than storage, data requests are made with smaller, more efficient 64-byte accesses. Each load or store instruction fetches a 64-byte cache line with low latency. At the Intel® Optane™ DC persistent memory module, this results in a read or write of 256 bytes of data accessed—much smaller than the 4 KB accessed in an Intel® Optane™ DC SSD.2 3 This means you can access data with smaller granularity and less overhead, and software developers can work with smaller structures to streamline their applications.
Intel® Optane™ Technology Performance for High-Throughput Uses
While the average-idle-latency advantages of Intel® Optane™ technology are impressive, real applications access data at high rates with concurrent reads and writes. In the presence of this type of heavy traffic, these applications especially appreciate low-latency accesses. To understand the performance advantage of Intel® Optane™ DC persistent memory, we return to the latency at load plots used to analyze the diﬀerences between SSDs. To simplify the discussion, we'll focus now only on the average latencies. Figure 5 shows just such latency for diﬀerent system throughputs to a single SSD or memory module. The three technologies pictured, with latencies progressively lower by 10x or more at any throughput, oﬀer a compelling storage hierarchy upon which applications can excel.
Intel® Optane™ DC persistent memory oﬀers by far the lowest latencies at all bandwidths for a truly unique performance with persistence.1 2 3 The very low average latency (blue line) oﬀered by the Intel® Optane™ DC persistent memory represents a truly new class of storage. This latency was already stated as between 100ns and 340ns depending on access locality. From the chart it's clear that the latency measured falls within the lower part of that range. That's because we measured 256 byte accesses made up of four cache lines, so while one cache line was amiss, the other three are hits, for an average latency at the lower fraction of the range. Notice that this low latency is consistent as the load is increased substantially. Even a single memory module can sustain throughputs significantly higher than either SSD tested.
Putting Persistent Memory to Use in the Data Center
There are many practical applications for Intel® Optane™ DC persistent memory because it oﬀers consistent low-latency, non-volatility, and high capacities. That makes it ideal for in-memory database applications that need mission-critical performance for large data volumes. For example, to remain competitive, financial services institutions rely on constantly changing trading data correlated with massive amounts of historical data. To process information quickly, these businesses frequently employ in-memory databases, using DRAM to provide real-time insights. But using DRAM has several drawbacks, including:
- Higher per-gigabyte costs compared to Intel® Optane™ DC persistent memory
- Limited capacity per DIMM
- Volatility, which means the massive dataset must be loaded into memory after every restart or power failure before the system is again available
In contrast, Intel® Optane™ DC persistent memory oﬀers:
- Lower cost per gigabyte, enabling an expanded data size within the same budget, or depending on the application a similar performance at lower cost
- Much higher capacity memory, with up to 512 GB DRAM per module
- Persistence for rapid recovery times after a power outage or system reset
With the larger capacity it makes available, Intel® Optane™ DC persistent memory can dramatically change the data-tiering landscape used by financial-services institutions and large enterprise businesses. Instead of migrating warm data from less costly NAND SSDs into DRAM as needed, organizations can create a large, non-volatile hot data tier in memory, giving data scientists and analysts real-time or near-real-time access to critical insights over that larger data set.
Intel® Optane™ technology is the first new memory technology to be produced in volume in decades. It has entered the platform as a solid state drive, accessible through the standard OS storage stack and APIs. As an Intel® Optane™ SSD, it delivers latency an order of magnitude lower than NAND SSDs and an even greater quality of service advantage. This new memory technology is also now available as system memory as part of an optimized system paired with 2nd Generation Intel® Xeon® Scalable processors. It is accessible directly from the applications without operating system overhead. Intel® Optane™ DC persistent memory delivers bigger capacity than DRAM and persistence at extremely low latencies. These technologies, with latencies progressively lower by 10x or more at any throughput, oﬀer a compelling storage hierarchy upon which applications can excel. Together, NAND SSDs, Intel® Optane™ SSDs and Intel® Optane™ persistent memory fill out the memory and storage hierarchy enabling data to be placed at the right level so that it may be accessed quickly to deliver excellent system performance.
Intel Fellow Frank Hady
Frank Hady is an Intel Fellow and the Chief Optane Systems Architect in Intel's Non-Volatile Memory Solutions Group (NSG). Frank leads research and definition of Intel® Optane™ technology products and their integration into the computing system. Frank has served as Intel's lead platform I/O architect, delivered research foundational to Intel® QuickAssist Technology (Intel® QAT), and driven significant platform performance advances. He has authored or co-authored more than 30 published papers on topics related to networking, storage, and I/O innovation and presents often on memory and storage. He holds more than 30 U.S. patents. Frank received his bachelor's and master's degrees in electrical engineering from the University of Virginia, and his Ph.D. in electrical engineering from the University of Maryland.