|
[Sponsors] |
January 20, 2015, 17:46 |
Optimal machine for pre/post/head-node
|
#1 |
Member
Kim Bindesbøll Andersen
Join Date: Oct 2010
Location: Aalborg, Denmark
Posts: 39
Rep Power: 15 |
I need some advice for optimal specifications for a new head node in my cluster. The head node will be running pre/post processing and storage, but not solving.
My present setup is a 32 core, 2 node cluster: Each node: Dual Xeon E5-2667 v2, octo-core. RAM: 64 GB, 1866 MHz Disk: 150 GB, 15k, RAID1 Infiniband interconnect. One node is assigned as head node and has additionally 1 TB, 10k, RAID5 for storage of all CFD files. Software: SpaceClaim ANSYS CFX, ANSYS Meshing, ANSYS CFX Post. Presently pre/post/meshing is done on local machine Xeon X5687, quadcore, 24 GB RAM 1333 MHz, 250 GB SSD. Ethernet to cluster at 1 Gbps. IT department wants to move cluster to remote location with low bandwidth to my desk (250 Mbps shared with 300 employees). So I see only solution is to also move pre/post/meshing to remote location and run accelerated graphics (Turbo VNC or HP Remote) to my desktop PC to avoid heavy net-traffic and long waiting times. In this setup all files will be at remote location, and only traffic will be keyboard, mouse and screen signal back. Any other suggestions to run CFD at low remote bandwidth? My plan is to add a new head node to the solver cluster, and run pre/post/meshing remotely on this node. As far as I understand single core speed is important as most meshing and post processes are single threaded. If meshing is parallel it will only be for partitioned meshes. How important is memory bandwidth? Is number of cores important? What does the task as head node demand? I know it takes som read/write at opening and storing files. But during solving I gues the only head node activity is storing transient results (if any). Does hyperthreading do any good for this type of tasks? Suggested specification for new head/pre/post/meshing node: i7-4790K, quad core, 4.0 GHz. Single core PassMark: 2534 Total PassMark: 11,246 Memory lanes: 2, DDR3-1600 Max memory bandwidth: 25.6 GB/s Max memory: 32 GB Graphichs: Disk for Windows Server and ANSYS software: 250 GB SSD (from current desktop workstation). Disk for CFD files storage: 1 TB, 10k, RAID5 (from current cluster head node). Alternative: i7-5930K, six core, 3.5 GHz. Single core PassMark: 2072 Total PassMark: 13,650 Memory lanes: 4, DDR4-2133 Max memory bandwidth: 68 GB/s Max memory: 64 GB (rest will be the same) As the head node does no solving I guess the latency of interconnect to solver node is not important, so a 10 Gbps Ethernet will do? (thus I avoid buying a Infiniband switch, which is very expensive). Any good advice and experience on configuration of pre/post/meshing/head node? Best regard Kim |
|
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
[TUTORIAL] Run fluent on distributed memory with 2 windows 7 64 bit machines | ghost82 | FLUENT | 54 | February 9, 2022 03:32 |
[ICEM] Error in mesh writing | helios | ANSYS Meshing & Geometry | 21 | August 19, 2021 14:18 |
Cluster ID's not contiguous in compute-nodes domain. ??? | Shogan | FLUENT | 1 | May 28, 2014 15:03 |
The fluent stopped and errors with "Emergency: received SIGHUP signal" | yuyuxuan | FLUENT | 0 | December 3, 2013 22:56 |
999999 (../../src/mpsystem.c@1123):mpt_read: failed:errno = 11 | UDS_rambler | FLUENT | 2 | November 22, 2011 09:46 |