June 12, 2013
It’s been just a tick under a year since Intel’s acquisition of Whamcloud and its Lustre braintrust. While it was a rather quiet transition, the chipmaker made some noteworthy strides toward pushing Lustre further into the mainstream--without ignoring its roots in HPC.
The company kicked two announcements about Lustre into the arena with high hopes of appealing to the enterprise crowd. Both news items are meant to strike a chord with the big business crowd--a market that’s increasingly interested in (you guessed it) “big data” and Hadoop, as well as general usability for ever-mounting scalability troubles.
Brent Gorda, founder and former CEO of Whamcloud, who now manages all things Lustre for Intel, said that the file system has a dangerous reputation as being hard to handle, despite its recognized benefits. This bad rap, however, spins from a misconception that’s fed from the stream of news out of multi-petabyte installations where Lustre really can require some serious expertise, according to Gorda.
Still, businesses are looking to plug in Lustre, at least according to Intel, which tends to pay attention to customer demand. While they want Lustre benefits, their hangup is that they don’t want to make new hires to run the file system. Hence the rollout today of Intel Enterprise Edition for Lustre, which comes with a new management tool called Intel Manager for Lustre.
As Gorda explains, it’s all about opening access to Lustre for the “rest of the world.” He says that currently, it’s the big national labs that are hiring people to run it since they need the scalability and reliability it offers. What Intel decided, based on customer demand, is that they want to offer approachable tools for Lustre to anyone who can manage a Linux box. The key here, argues Gorda, is not just the benefits of the management interface-- it’s the Lustre support that’s now backed by a tier-1 vendor (versus little old Whamcloud peddling its support that made the big shops shy).
As with any software project, this mainstreaming of Lustre will be a process of evolution. But at this point, the management tool simplifies interacting with Lustre so users can wick away the need to deal with command lining, finding configuration files all over the place and are freed from watching logs scroll by. The management tool does all of this, thus Intel argues, opening access to Lustre without making companies add more bodies to deal with it--making it far more attractive to some key areas that are scaling at HPC levels (oil and gas, financial services...the usual suspects).
In addition to the management tool to facilitate wider adoption of Lustre in new environments, the company made a Hadoop announcement set to sing to those frustrated with the native Hadoop file system (HDFS), at least for those who have tapped Intel's distro. Gorda described how one talented Lustre and HDFS expert steamrolled Hadoop with a new Java class to fully swap in Lustre over HDFS. The impetus here was that many of the traditional HPC sites were talking to Intel about their Hadoop plans. Of course, the needs of a file system for Hadoop at the Web 2.0 scale wouldn’t necessitate such a swap-out since it runs on local disk for most. For the HPC’ers, however, the disks were probably lifted from the boxes a decade ago, so the challenge is convincing Hadoop to run off a global parallel file system--a trick he says they’ve pulled off nicely.
Further, Gorda said that in a conversation with an oil and gas company about the pending HDFS news, the business did the math and found that Hadoop’s triple-replication was going to cost them quite a bit more than Lustre.
When it comes to Hadoop, “Lustre is performing faster than local disk because we’re feeding you data at network bandwidths, which are generally faster. When it comes time for the sort/merge stage, you already have a global scratch space, which means you don’t have to do all the communication steps that are necessary with local disk,” Gorda explained.
During our chat yesterday, Gorda seemed genuinely enthusiastic about how Intel’s might has led to increasing adoption and interest in Lustre. He said that prior to coming Intel’s wing, they would talk to potential big name customers in financial services and other areas who were in need of a robust file system for large node-counts but were put off by the fact that Whamcloud had only a small team. With Intel’s namepower and funding--both through the Fast Forward program and the company’s own ambition to deliver Lustre to new markets based on demand--Gorda says Lustre will continue to meet the eventual needs of exascale and the current demands from enterprise users who have smacked into the scaling wall with their current file systems.
Outside of today’s news items around Lustre, Gorda said there are some cool things brewing for the coming months, particularly powered by Fast Forward program funds.
For instance, he pointed to work being done on so-called burst buffers and new innovations fed by solid state. . As one can imagine, a file system that’s looking toward such technologies might be laden with some interesting features that could dramatically speed data and focus on latency as the data moves ever closer to the compute.
He also talked about how his team is adding an object storage model to Lustre. As he noted, “POSIX is getting long in the tooth; it was not designed to go this far due to locking and semantic issues. We’re going down the path of building out--specifically for HPC--an object-based interface that is different than today’s cloud object models. The goal is to offer a way to communicate directly to the application about where its data is self-consistent. So if the application is humming along, clicking in at checkpoints, you’ll know, for example, that your fifth checkpoint is consistent when all our tasks check in and move on to the next iteration. Intel’s Lustre team is refining this development currently.
As for the future of Lustre in environments outside of the traditional purview of HPC, Gorda remains confident that the power of Intel will be enough to add some new sway. The new developments to add to usability help boost Lustre’s profile and once word hits the mainstream Hadoop-obsessed press with news of a super-scalable HDFS replacement, it could mean big things for the file system’s future.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?