Although there wasn’t a lot of earth-shattering news out of SC16 this year – only a few major product announcements in what used to be a central part of the show – there was a lot of evidence of real progress in technologies needed to support next-generation and Exascale computing infrastructure. The major “miracle-needed-here” roadblocks we have seen in years past in at least one of compute, networking, or storage technologies were not really present (or at least not driving the discussion) this year.
FPGAs and code modernization were big topics, as usual, but with relatively no big news. FPGAs continue to be a credible threat to traditional microprocessors, and the perceived lack of innovation of market leaders is widely thought once again to be opening the door to challengers like ARM.
The big server vendor news was definitely HPE’s purchase of SGI, which boosted booth traffic for both SGI and HPE, and they capitalized on that quite nicely.
The big networking news was the rapid adoption of Intel Omni-Path and Mellanox’s (re)announcement of their 200Gbit Switches. An interesting point to note here is not that the High Data Rate (HDR) InfiniBand and Ethernet solutions that Mellanox is building have a theoretical bandwidth of 200 Gb/s, but that they purport to deliver that rate at only 1% of CPU utilization.
Omni-Path as a server interconnect was, if not Omni-present, at least very much in evidence in a number of places in the conference. As a direct connection to storage, however, it is still a narrow field, with DDN in the lead. You can listen to JCAHPC talk about their Omni-path success here.
For storage this year, the big topics were the transition to SSDs and the rise of burst buffers.
With SSD supplies tightening, $/GB is not dropping as fast as it usually does for new storage media, and there were plenty of questions about how supply would affect the transition. More common, however, were conversations about how best to take optimal advantage of flash – whether an all-flash array or flash as a software-defined tier is the answer to accelerate metadata, data, or applications.
Several vendors touted solutions under the burst buffer terminology, but only a few (DDN IME ® and Cray DataWarp) appear to have advanced past the science project or slideware phase. DDN’s Vendor Forum talk on IME was standing room only – unheard of for a vendor forum presentation! JCAHPC’s new system, Oakforest-PACS, powered by a DDN Infinite Memory Engine (IME) Burst Buffer rated at 1.56 TB/scame in at #6 on the Top500, using 1/10 to 1/100 the amount of gear other TB/s+ file system sites need. This is a MAJOR milestone in HPC, and DDN is proud to be part of it. DDN’s heavily-trafficked IME demo illustrated our new commodity-based burst buffer nodes, demonstrating advanced data protection, performance under a variety of I/O scenarios, and, of course, peak performance.
DDN broke all our previous records for SC – customer meetings, user group attendees, booth visitors, and party attendees – and we also won a number of new HPCWire Awards, including best product. All in all, a pretty fabulous show for us!