HPC DIRECTOR
The ins and outs of HPC
Lee Ward, principal member of technical staff at Sandia National Laboratories, tackles HPC-related IO improvements to multiple supercomputing projects
rimarily, as a researcher in high performance IO and file systems, I’ve contributed file systems and IO improvements to multiple supercomputing projects at Sandia. Probably the most visible was the virtual file system interface for our Red Storm machine, the prototype for Cray’s XT line.
P
Sandia runs a number of scientific machines. Our largest, most popular, machines are: l Red Sky, which has 5,300 nodes, eight cores per node (42,400 processors) running Centos 5.3 Linux. This machine is currently at number 10 on the Top500 list; l Thunderbird, which has 4,500 nodes, two cores per node (9,000 processors) running Redhat Enterprise Linux; l TLCC (two identical machines), featuring 576 nodes, 16 cores per node (9,216 processors) in the Tri-Lab Linux Capacity Cluster (TLCC) running the tri-lab TOS software stack, a derivative of Redhat Enterprise Linux; and l Red Storm, which was the prototype for Cray’s XT line of machines. The machine was developed in a partnership with Cray, beginning in the early part of the last decade. The machine has been upgraded twice since. Currently, it has 38,208 processors, and runs Catamount, a Sandia-crafted, special-purpose lightweight operating system designed to maximise the capabilities of large multi-program parallel machines.
Sandia is a multi-programme laboratory. Primarily, we work on the United States national security mission. Together with
30
Los Alamos National Laboratory and Lawrence Livermore National Labs, we work on maintaining this nation’s nuclear weapons stockpile.
Major areas that we are working on within that mission include life extension programmes for the weapons, reliability, security, and safety. We are not solely limited to that, though. Sandia’s mission truly is ‘exceptional service in the national interest’ and we have made significant contributions in nuclear non-proliferation, homeland security technologies, energy research, synthetic aperture radar, robotics, and more.
‘We are well into the direction where we try to build a centralised store and connect the machines to it’
Our machines’ typical use reflects these missions. Sandia is fundamentally an engineering laboratory, so we would be modelling many things one might expect, such as heat transfer, shock and materials deformation. Obviously, access to these machines for our sensitive work takes place in restricted areas, usually local to the facility. We do much that is not sensitive, though, and access can be gained through the internet – much like any large business. Here, we use the normal technologies; firewalls, gateways, VPN, encrypted HTTP and so on.
SCIENTIFIC COMPUTING WORLD APRIL/MAY 2010
I’m an IO guy, so I’d like to elaborate
on our IO solutions. We see the world evolving in the high-performance space. In the past, these large machines have incorporated dedicated IO systems. The storage facilities were directly attached and not shared with our other compute resources.
At this point, we are well into a direction where we try to build a centralised store and connect the machines to this shared resource. Our first attempt at this was a purchase from Panasas. We bought 10 shelves of storage a number of years ago. Our special needs for this purchase led us to a non-standard configuration.
Since we primarily use the resource for parallel IO access from the supercomputers, we have only one DirectorBlade for every two shelves, instead of the recommended one per shelf. Our second solution leverages two
models of Direct Data Networks (DDN) RAID controllers (9900s and 9950s). The software used to aggregate these is Sun’s Lustre file system. This store was assembled over multiple purchases and is about one petabyte of capacity currently. These two stores service all of the
machines described above, except Red Storm.
Red Storm represents what is probably the last of the old-style, dedicated IO system architecture. That machine has a mixture of DDN and LSI RAIDS and aggregates it with the Lustre file system software.
www.scientific-computing.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44