3Leaf Makes x86 Boxes into SMP Cloudware
Claims it will revolutionize data center cost models
Open Source and Cloud Computing on Ulitzer
3Leaf Systems, the well-funded start-up, dropped its fig leaf Tuesday and took a running jump into the pools of memory, I/O and cache that it can construct and deconstruct at will based on the application, creating scale-up shared-memory SMP systems the likes of mainframes, proprietary mid-range machines and pricey RISC-based Unix boxes out of racks of cheap and dirty two-socket scale-out x64 systems.
Lately things have been going mostly the other way around.
3Leaf claims it will revolutionize data center cost models.
Gabriel Consulting principal Dan Olds thinks it's "definitely a technology to watch."
He says, "3Leaf Systems is pioneering a unique hardware approach that will allow a single operating system instance to expand across multiple server blades, boxes or racks; or shrink to just a fraction of a single server. This will give customers the ability to scale various system resources to match a given workload, increasing overall utilization and efficiency."
Ironically the Intel-backed 3Leaf has started building its so-called Dynamic Data Center (DDC) out of AMD parts with first server deliveries slated for next month.
What the start-up's got is an ASIC and some software that can merge the isolated CPU, I/O and memory resources of multiple x86 systems into one big computing "fabric" so resources can be instantly delivered, just-in-time, to demanding applications for a lot less money than it normally takes.
Currently, using other people's widgetry, if a virtual machine starts to peak and has to go to another two-CPU server it can't take the resources it was using with it. It has to start all over again. That's a bummer 3Leaf means to overcome.
3Leaf's so-called DDC-ASIC is supposed to crunch through complex coherency protocols at lightning speed and enable distributed cache coherency among all system cores - complements of layering on top of 20 Gbps Infiniband or 10 Gbps Ethernet - creating the basic control structure needed for agility.
It's available now for AMD's quad-core Shanghai and six-core Istanbul Opteron processors supporting AMD's HyperTransport interconnect and up to 1TB of main memory across up to 16 nodes. Figure support for AMD's expected 12-core Magny Cours chip in Q1.
3Leaf's next version will support 32 nodes and up to 64TB of main memory using Intel CPUs compatible with the Intel Quick Path Interconnect (QPI) 1.1 beginning with the promised Sandy Bridge processors. Figure the end of next year or 2011.
AMD's HT-supporting widgetry was out first, which is why AMD got precedence. 3Leaf extends the HT bus to switches and virtualizes the chip's memory and cores.
CEO BV Jagadeesh claims 3Leaf's timing is impeccable. It needed the virtualization support and memory management tables now built into x86 chips, operating systems that can add and subtract resources, and low-latency commodity switches. It's a way of brushing off the start-up's aborted venture into I/O virtualization.
Anyway, 3Leaf's can currently provide an Opteron-based Dynamic Data Center Server (DDC-Server) made by Super Micro fitted out with its widgetry. The box is OEM-ready and up for adoption for cloud computing, data warehousing and data mining, HPC and social media.
3Leaf doesn't want to sell it itself; it wants OEMs to do it, but 3Leaf is making certain fixed configurations available to end users.
It figures that by integrating its technologies, OEMs can expand their offerings, tap into new market opportunities or fill a current product gap. In the process, they can get significantly lower costs, faster time-to-market, and the possibility of serving both the Intel and AMD x86 markets.
Users can treat these x86 servers as building blocks to create dynamic "systems" of any configuration, with the ability to push and pull specified resources across silos on-the-fly ultimately without rebooting.
3Leaf's technologies should also enable massively scalable cloud infrastructures based on these cost-optimized building blocks.
A DDC-Server configured with 1TB of shared memory, 192 2.8GHz Istanbul cores and 8TB of storage, all connected via an InfiniBand switch, with Linux and 3Leaf's DDC-Pool software lists for $250,000. One with 256GB of shared memory, 96 2.4GHz Istanbul cores and 4TB of storage runs $99,000. Each two-node server needs an ASIC.
The widgetry can use SSD, FC or SATA storage and 10GbE instead of InfiniBand.
3Leaf imagines in-memory databases that run thousands of times faster than we're used to. Large social media analysis that can find a needle in a haystack, transaction processing systems that don't stall, and HPC apps that are more accurate. It also imagines avoiding wasted memory because of unnecessary data replication and consolidating unused capacity into large enough usable capacity.
Its boxes should reduce opex by minimizing operating system administration, power and cooling, and capex by using garden-variety two-socket servers.
3Leaf says its DDC-Server provides hardware partitioning for fault isolation at the board/blade level and logical fault isolation so one OS instance can't bring down other instances - and it can offer high-availability failover across two or more instances of an OS within or across server racks.
3Leaf's software stack includes DDC-Pool, which coalesces multiple x86 boxes or blades into a single larger contiguous system, and supports the static reconfiguration of an OS across whole servers or blades in a compute cluster when rebooted; DDC-Share, which provides a greater degree of flexibility, allocating resources down to the core level and allowing instances of an OS to run across whole or parts of servers or blades, with static reconfiguration of the OS at reboot; and sometime next year DDC-Flex, which is supposed to provides runtime reconfiguration of OS images across any portion of a compute cluster. An OS image can expand or shrink in terms of CPU, memory and I/O resources without a reboot.
Pool and Share, both available now, work with Linux. Windows support is scheduled to be available next year.
In August, 3Leaf Systems broke the previous SPECjbb2005 performance record set by x86 systems by a factor of two. That's the current benchmark for eCommerce transaction processing and evaluates the performance of server-side Java applications.
3Leaf is focused initially on HPC which faces challenges in flexibility, cost management, software optimization time and realized performance as a fraction of peak speeds. Its roadmap also targets analytics, data warehousing and BI, the e-commerce side of OLTP and enterprise OLTP.
It's not good for everything; only apps that can deal with a NUMA architecture can run unmodified.
3Leaf is backed by Alloy Ventures, Enterprise Partners, Intel Capital, LSI and Storm Ventures to the tune of $67 million so far. The company is expected to go out for a D round next year. LSI, which led its C round, makes its ASICs.
