Hu Yoshida

Fake News Around FMD versus SSD Performance and Availability

Blog Post created by Hu Yoshida on Mar 31, 2017

fake news.jpg

 

Competitors who lack the engineering capability to design their own flash devices use standard SSDs that were designed for the high volume commodity PC and server markets. These competitors are trying to create FUD around our purpose built enterprise flash module, FMD, claiming that the offload process causes significant backplane issues and the loss of an FMD would impact performance and cause other issues due to the design of our offering. This is utter nonsense and comes under the category of “fake news

 

Hitachi has taken a very different approach to flash storage devices. Unlike other flash array vendors that use standard SSD’s, Hitachi has built their own flash module (FMD) from scratch to integrate the best available flash controller technology into the storage portfolio. While Hitachi does support SSDs, they recognized the opportunity to deliver higher capacity flash drives with advanced performance, resiliency, and offload capabilities beyond what SSDs provide by developing our own flash module, with a custom flash controller.

 

Industry Analyst, George Crump, stated in one of his blogs: "With flash memory becoming the standard in enterprise SSD systems, users need to look more closely at the flash controller architectures of these products as a way to evaluate them. The flash controller manages a number of functions specific to this technology that are central to data integrity and overall read/write operations. Aside from system reliability, poor controller design can impact throughput, latency and IOPS more than any other system component. Given the importance of performance to SSD systems, flash controller functionality should be a primary focus when comparing different manufacturers’ systems"

 

Gartner’s August 2016 report (ID G00299673) on Critical Capabilities for Solid-State Arrays recognized this as a trend “To gain increased density and performance, an increasing number of vendors (e.g., Hitachi Data Systems, IBM, Pure Storage and Violin Memory) have created their own flash NAND boards, instead of using industry-standard solid-state drives (SSDs). Dedicated hardware engineering has reappeared as a differentiator to industry-standard components, moving away from the previous decade's trend of compressed differentiation”. Hitachi Data Systems started this trend in 2012 with the announcement of our first FMD.

 

Competitors who lack the engineering capability to design their own flash devices use standard SSDs that were designed for the high volume commodity PC and server markets. These competitors are trying to fight this trend by creating FUD around our FMD, claiming that the offload process causes significant backplane issues and the loss of an FMD would impact performance and cause other issues due to the design of our offering. This is utter nonsense and comes under the category of “fake news

 

In the first place, the only function that is off loaded from the storage controller to the FMD is data compression which is handled by two coprocessors in the FMD and has no impact on performance. Compare this to the software overhead of doing the compression/decompression for selected SSDs in the storage array controller versus doing this in hardware in the FMD. Because of the performance impact of doing compression in the storage array controllers, storage administrators have the added task of managing  the use of compression. With the FMD you can turn on the compression and forget it. Aside from the reporting of audit log and normal usage information, there is no significant side band communication between the Hitachi VSP storage controller and the FMDs so the claim that the offload process causes significant backplane issues is completely false.

 

The management of a flash device is very compute and bandwidth intensive. SSDs rely on a controller to manage the performance, resilience and durability of the flash device. As the I/O activity increases, these functions cause controller bottlenecks, sporadic response times, and a poor customer experience, causing IT organizations increasing work in managing the tradeoffs between workloads and data placement on different SSDs. We have published a white paper that explains the technology, and the pitfalls of trying to do this with the limited architecture and compute power of standard SSDs. You can down load this white paper via this link.

 

FMD Controller.png

 

The brains of the FMD is a custom-designed ASIC featuring a quad core processor, 32 parallel paths to the flash memory, and 8 lanes of PCIe v2.0 connection to the external SAS target mode controllers. This ASIC is composed of more than 60 million gates, two coprocessors for compression and decompression, and direct memory access (DMA) assist. The ASIC compute power drives up to eight times more channels and 16 times more NAND packages than typical 2.5 inch SSDs. This powerful ASIC enables the FMD to avoid the limited capabilities of the standard SSD controllers, which restricts the amount of flash that an SSD drive can manage and forces the flash array controller and storage administrator to do more management in the placement of data on SSD drives. The FMD ASIC enables Hitachi Data Systems to deliver capacities of up to 14 TB in one FMD today.

 

Unlike standard SSDs, the FMD was purpose built for enterprise storage workloads. It was specifically designed to address intensive large-block random write I/O and streams of sequential write requests from applications such as software as a service, large scale transaction processing, online transaction processing (OLTP) databases, and online analytic processing(OLAP). It also understands that storage array controllers format drives by writing zeros, so the FMD avoids writing the zeros, to improve performance and durability. It can also erase all the cells, even the spare cells that the array controller cannot see and report back the status of the erased cells through a SCSI read long command for auditing purposes. SSD storage arrays have no way to securely erase all the flash cells in an SSD since they cannot see the spare cells and overwrites are always done to a new cell.

 

As far as the loss of an FMD having an impact on performance there are two cases, a planned threshold copy-replacement and an unplanned RAID reconstruction. With the copy-replacement, the simple copy places a minor load on the FMD, but there is no impact to host I/O. There is an impact for the standard RAID reconstruction, but that is the same for any storage array, except that higher performance of the FMD could shorten the reconstruction time, depending on what else is happening in the array.

 

To get the true news on Hitachi Data System’s FMD please read our white paper which explains it all.

Outcomes