Tag Info

Hot answers tagged

10

We did this in the high-frequency financial trading world for a short while (close parallel to HPC, given the application architecture I was working with)... Around early 2010, I was deploying to custom 3U rack-chassied, single socket i7 "gaming systems" with 10GbE Solarflare (with OpenOnload UDP message kernel-bypass) and/or Infiniband interconnects. I ...


6

The i7 can't use ECC Buffered RAM and can't be installed in dual or quad socket configurations. That seems like reason enough to not use it - but of course, your needs might dictate otherwise.


5

You should distinct between three types of technologies: 1) OS Kernel mods (in this case, modules or kernel patches) [This technology is software-based] LinuxPMI/OpenMosix (http://linuxpmi.org/) OpenSSI (http://openssi.org/) Kerrighed (http://www.kerrighed.org/) In the technologies listed above, you modify the OS to provide the user a "feel" of an SMP, ...


4

Scalability, for one. Also, the fact that it is open source is important, since super computers exist almost solely for research and researchers love the flexibility of open-source software. They can hack and slash the OS to be ultra-tuned to fit their very specific need at the small cost of the source code and some bandwidth for anyone that's curious about ...


4

So there is more than one flavor of cluster out there, RH Cluster provides just one... HPC the other. A quick from-memory run down: High Availability/Active-Passive. Sometimes referred to as N+1 (or N+M) clusters. In general, a process or an application runs on a single server. Then, via shared storage, moved IP addresses, and some other cluster magic, ...


4

Apologies for answering this myself: we encountered the answer just a minute ago. It turns out that if you call slmgr with cscript, the output will go to the command line rather than a dialog box. So you want a command like, cscript slmgr c:\windows\system32\slmgr.vbs -dli


4

NFS is from 1 client to 1 server, so the overall performance is limited by the performance of that 1 server. Adding more servers does not help. Lustre splits the data, the data gets requested from 1 server, but can be sent from one or more other servers. So adding more servers does help (which is why "Lustre scales"). This is an important bit from your ...


3

In OpenFabrics, the infiniband-diags package comes with a tool called "ibroute" that should do exactly what you're asking for. There is a script called "dump_lfts.sh" in infiniband-diags that wraps around ibroute and will output the routing tables for all the switches in your fabric. There are other scripts out there to wrap around these for further ...


3

No, AD LDS does not have the concept of member servers or computers joining a domain. It's just an LDAP front-end. You need Active Directory Domain Services.


3

Well, to be frank I went the simple route and happily used iSCSI over IP over IB and it worked easily and performed well: Infiniband IP setup ultra-quick primer. first... install opensm, infiniband-diags, rds-tools, sdpnetstat, srptools, perftest (for benchmarks) load IB driver module, ib_umad, ib_ipoib now you have a new network interface to configure. ...


2

I recently configured SRP target/initiator pair on Linux, and had ~100% performance increase (580MB/s on 10Gbps SDR) over traditional iSCSI-over-IPoIB configuration (300MB/s on SDR). Setup: Distribution: Debian sid Linux kernel: 3.4.0-rc1 (3.3 or above is required for in-kernel SRP) Infiniband stack: OFED-1.4 (which comes with Debian) SRP/iSCSI target: ...


2

The idea of a Supercomputer isn't one box with all the resources as that would be difficult to do. The idea is instead to share the load across hundreds/thousands of nodes with Hi-speed networks for transfer between them.


2

You should consider Condor and one of the forks of Sun Grid Engine. Both Condor and SGE are heavily used in the academic HPC community for batch scheduling, and will allow you to scale gracefully should you acquire additional hardware.


2

Most of the parallel I/O benchmarks on that page should actually write data (albeit often random noise). I've personally only used IOR and PIO-Bench from that list, but these are all relatively well-known and accepted benchmarks in the HPC community. "Relatively", because I/O benchmarks can produce holy wars in HPC (especially vs vendors)... but none of ...


2

It's a crapshoot. There's a certain amount of spin-up/spin-down latency associated with changing the CPU frequency, and this can negatively affect certain workloads (typically ones that are CPU-hungry but still memory-bound, resulting in uneven but high-throughput CPU access patterns). Modern CPUs seem to perform much better with this than even just two ...


2

This is a major source of problems in the Hadoop ecosystem and AFAIK there isn't a good answer to comprehensive planning for this kind of resource. Overall, the isn't an Enterprise quality Hadoop distribution that will support the laudable level of diligence that you are applying to your system. I am pretty sure that there will be one in the next few ...


2

I don't think that you can do what you want with OpenMP. OpenMP is designed for shared memory programming. One example of this is using multiple cores and/or processors on the same machine. A cluster (at least the typical clusters made from commodity hardware) are distributed memory. Commonly MPI is used to program these types of machines. There are ...


2

Wow... earned tumbleweed badge on this one! The answer, if anyone else ever needs it is qacct -j 221466 where 221466 is the job id


2

A short sketch: The most important part will be planing of your networks and the physical installation. You'll probably use at least three or four networks. One for management where you provision, monitor, deploy, configure your nodes, one for storage access, and another one for things like MPI and internode communication and an out-of-band network where ...


2

There is not a lot of easily discoverable technical information available for the Phi yet. (The scant Intel pile, including the Software Developer/User guides). All the press suggests that the Phi is meant to run along-side newer generation Xeons (E5-2600 and 4600 lines specifically), but isn't telling us what the architectural reason for this is. Intel's ...


2

Personally I would still lean towards the Xeon as it is designed to support more concurrent load over a longer period of time. If you are running CPU intensive work over an extended period of time, say hours or weeks the Xeon has a much longer mean time to failure than the i7 does. Other than that I defer to @MDMarra's response as to server load out.


2

If you use xfs, you will always have that problem: (quotaon manpage) "XFS filesystems are a special case - XFS considers quota information as filesystem metadata and uses journaling to provide a higher level guarantee of consistency. There are two components to the XFS disk quota system: accounting and limit enforcement. XFS ...


2

This is most likely not a limitation of NFS you are encountering here. Also take into account that those 5 GBytes take at the very, very least 40s to transfer at gigabit wire speed - for each client. You have 32 of them hammering the head2, and they're not likely to request the same blocks at the same time. Add Ethernet, TCP/UDP and NFS overhead, and you'll ...


2

Since you are doing performance analysis, the first question should be: "what is the data I am basing the assumption on? Are there network traces or other performance data that would support this hypothesis?" There are a lot of possible bottlenecks in such a system, and I would question the choice of the network filesystem last, especially since you do not ...


2

The goal is to get it so that every layer between the old storage and new storage goes faster than the maximum read speed you can get from your old machine. Their specs claim 6GB/s sequential (which this should be). That means that the minimum time possible to move the data would be in the realm of 46 hours, if you are able to get the advertised speed. ...


1

OpenStack does not provide any support on its own for distributed shared memory.


1

vSMP Foundation supports for "nested" virtualization. As such, you could run KVM within a VM that is aggregated from multiple servers, using vSMP Foundation. ScaleMP announced that in the future it will be possible to run vSMP Foundation in a "nested" VM - so vSMP Foundation is running on multiple nodes, each one running KVM, and vSMP Foundation inside it. ...


1

Probably, yes. It's hard to tell which will be faster though because there are a lot of factors. If possible, run a smaller version of your program so that you can compare both. Here are some factors that affect performance: Interprocess communication (as Michael mentioned in his comment): Communication between cores on the same node is faster than ...



Only top voted, non community-wiki answers of a minimum length are eligible