1. Using Kure and Topsail Mark Reed Grant Murphy Charles Davis ITS Research Computing

2. 2  Compute Clusters Topsail Kure  Logging In  File Spaces  User Environment and Applications, Compiling  Job Management Outline

3. 3 Logistics  Course Format  Lab Exercises  Breaks  UNC Research Computing http://its.unc.edu/research  Getting started Topsail page http://help.unc.edu/6214  Getting started Kure page http://help.unc.edu/ccm3_015682

4. What is a compute cluster? What exactly is Topsail? Kure?

5. 5 What is a compute cluster? Some Typical Components  Compute Nodes  Interconnect  Shared File System  Software  Operating System (OS)  Job Scheduler/Manager  Mass Storage

6. 6 Compute Cluster Advantages  fast interconnect, tightly coupled  aggregated compute resources  large (scratch) file spaces  installed software base  scheduling and job management  high availability  data backup

7. 7 Initial Topsail Cluster  Initially: 1040 CPU Dell Linux Cluster 520 dual socket, single core nodes  Infiniband interconnect  Intended for capability research  Housed in ITS Franklin machine room  Fast and efficient for large computational jobs

8. 8 Topsail Upgrade 1  Topsail upgraded to 4,160 CPU replaced blades with dual socket, quad core  Intel Xeon 5345 (Clovertown) Processors Quad-Core with 8 CPU/node  Increased number of processors, but decreased individual processor speed (was 3.6 GHz, now 2.33)  Decreased energy usage and necessary resources for cooling system  Summary: slower clock speed, better memory bandwidth, less heat, quadrupled the core count Benchmarks tend to run at the same speed per core Topsail shows a net ~4X improvement Of course, this number is VERY application dependent

9. 9 Topsail – Upgraded blades  52 Chassis: Basis of node names Each holds 10 blades -> 520 blades total Nodes = cmp-chassis#-blade#  Old Compute Blades: Dell PowerEdge 1855 2 Single core Intel Xeon EMT64T 3.6 GHZ procs 800 Mhz FSB 2MB L2 Cache per socket Intel NetBurst MicroArchitecture  New Compute Blades: Dell PowerEdge 1955 2 Quad core Intel 2.33 GHz procs 1333 Mhz FSB 4MB L2 Cache per socket Intel Core 2 MicroArchitecture

10. 10 Topsail Upgrade 2  Most recent Topsail upgrade (Feb/Mar ‘09)  Refreshed much of the infrastructure  Improved IBRIX filesystem  Replaced and improved Infiniband cabling  Moved cluster to ITS-Manning building Better cooling and UPS

11. 11 Top 500 History  Top 500 lists comes out twice a year ISC conference in June SC conference in Nov  Topsail debuted at 74 in June 2006  Peaked at 25 in June 2007  Still in the Top 500

12. 12 Current Topsail Architecture  Login node: 8 CPU @ 2.3 GHz Intel EM64T, 12 GB memory  Compute nodes: 4,160 CPU @ 2.3 GHz Intel EM64T, 12 GB memory  Shared disk: 39TB IBRIX Parallel File System  Interconnect: Infiniband 4x SDR  64bit Linux Operating System

13. 13 Multi-Core Computing  Processor Structure on Topsail 500+ nodes 2 sockets/node 1 processor/socket 4 cores/processor (Quad-core) 8 cores/node  http://www.tomshardware.com/2006/12/06/quad-core-xeon-clovertown-rolls-into-dp-servers/page3.html

14. 14 Multi-Core Computing  The trend in High Performance Computing is towards multi-core or many core computing.  More cores at slower clock speeds for less heat  Now, dual and quad core processors are becoming common.  Soon 64+ core processors will be common And these may be heterogeneous!

15. 15 The Heat Problem Taken From: Jack Dongarra, UT

16. 16 More Parallelism Taken From: Jack Dongarra, UT

17. 17 Infiniband Connections  Connection comes in single (SDR), double (DDR), and quad data rates (QDR). Topsail is SDR.  Single data rate is 2.5 Gbit/s in each direction per link.  Links can be aggregated - 1x, 4x, 12x. Topsail is 4x.  Links use 8B/10B encoding —10 bits carry 8 bits of data — useful data transmission rate is four-fifths the raw rate. Thus single, double, and quad data rates carry 2, 4, or 8 Gbit/s respectively.  Data rate for Topsail is 8 GB/s (4x SDR).

18. 18 Topsail Network Topology

19. 19 Infiniband Benchmarks  Point-to-point (PTP) intranode communication on Topsail for various MPI send types  Peak bandwidth: 1288 MB/s  Minimum Latency (1-way): 3.6  s

20. 20 Infiniband Benchmarks  Scaled aggregate bandwidth for MPI Broadcast on Topsail  Note good scaling throughout the tested range (from 24-1536 cores)

21. 21 Kure  The newest, “latest and greatest” compute cluster in RC  Named after the beach in North Carolina  It’s pronounced like the Nobel prize winning physicist and chemist, Madame Curie

22. 22 Kure Compute Cluster  Heterogeneous Research Cluster  Hewlett Packard Blades  79 Compute Nodes, mostly Xeon 5560 2.8 GHz Nehalem Microarchitecture Dual socket, quad core 48 GB memory over 600 cores some higher memory nodes  Infiniband 4x QDR  priority usage for patrons Buy in is cheap  Storage Scratch space same as emerald No AFS home

23. 23 Kure Cont.  The current configuration of Kure is mostly homogeneous but it will become increasingly heterogeneous as patrons and others add to it.  Most login nodes are 48 GB but there are currently four high memory nodes  2 nodes each with 128 GB of memory  2 nodes each with 96 GB of memory

24. 24 Topsail/Kure Comparison Topsail  homogeneous  4000+ cores  2.33 GHz cores, Intel Core microarch.  12 GB memory/node  IB 4x SDR interconnect Kure  heterogeneous  600+ cores  2.8 Ghz cores, Intel Nehalem micorarch.  48 GB memory/node  IB 4x QDR interconnect

25. 25 Login to Topsail/Kure  Use ssh to connect: ssh topsail.unc.edu ssh kure.unc.edu  SSH Secure Shell with Windows see http://shareware.unc.edu/software.htmlhttp://shareware.unc.edu/software.html  For use with X-Windows Display: ssh –X topsail.unc.edu or ssh –X kure.unc.edu ssh –Y topsail.unc.edu or ssh –Y kure.unc.edu  Off-campus users (i.e. domains outside of unc.edu) must use VPN connection

26. File Spaces

27. 27 Topsail File Space  Home directories /ifs1/home/ anyone over 15 GB is not backed up  Scratch Space /ifs1/scr/ over 39 TB of scratch space run jobs with large output in this space  Mass Storage ~/ms

28. 28 Kure File Space  Home directories /nas02/home/ / /  a = first letter of onyen, b = second letter of onyen hard limit of 15 GB  Scratch Space – still evolving /nas – to be upgraded to 15 TB /largefs – to be upgraded to 30 TB run jobs with large output in these spaces  Mass Storage ~/ms

29. 29 Mass Storage “To infinity … and beyond” - Buzz Lightyear  long term archival storage  access via ~/ms  looks like ordinary disk file system – data is actually stored on tape  “limitless” capacity  data is backed up  For storage only, not a work directory (i.e. don’t run jobs from here)  if you have many small files, use tar or zip to create a single file for better performance  Sign up for this service on onyen.unc.edu

30. User Environment and Applications, Compiling Code Modules

31. 31 Modules  The user environment is managed by modules  Modules modify the user environment by modifying and adding environment variables such as PATH or LD_LIBRARY_PATH  Typically you set these once and leave them  Note there are two module settings, one for your current environment and one to take affect on your next login (e.g. batch jobs running on compute nodes)

32. 32 Common Module Commands  module avail module avail apps  module help  module list  module add  module rm Login version  module initlist  module initadd  module initrm More on modules see http://help.unc.edu/CCM3_006660

33. 33 Parallel Jobs with MPI  There are three implementations of the MPI standard installed: mvapich mvapich2 (currently only on topsail) openmpi  Performance is similar for all three, all three run on the IB fabric. Mvapich is the default. Openmpi and mvapich2 have more the the MPI-2 features implemented.

34. 34 Compiling MPI programs  Use the MPI wrappers to compile your program mpicc, mpiCC, mpif90, mpif77 the wrappers will find the appropriate include files and libraries and then invoke the actual compiler for example, mpicc will invoke either gcc or icc depending upon which module you have loaded

35. 35 Compiling on Topsail/Kure  Serial Programming Intel Compiler Suite for Fortran77, Fortran90, C and C++, - Recommended by Research Computing  icc, icpc, ifort GNU  gcc, g++, gfortran  Parallel Programming MPI (see previous page) OpenMP  Compiler tag: -openmp for Intel -fopenmp for GNU  Must set OMP_NUM_THREADS in submission script

36. 36 Debugging - Totalview  If you are debugging code there is a powerful commercial debugger, totalview  See http://help.unc.edu/CCM3_021717http://help.unc.edu/CCM3_021717  parallel and serial code  Fortran/C/C++  GUI for source level control  too many features to list!

37. Job Scheduling and Management

38. 38 What does a Job Scheduler and batch system do? Manage Resources  allocate user tasks to resource  monitor tasks  process control  manage input and output  report status, availability, etc  enforce usage policies

39. 39 Job Scheduling Systems  Allocates compute nodes to job submissions based on user priority, requested resources, execution time, etc.  Many types of schedulers Load Sharing Facility (LSF) – Used by Topsail/Kure IBM LoadLeveler Portable Batch System (PBS) Sun Grid Engine (SGE)

40. 40 LSF  All Research Computing clusters use LSF to do job scheduling and management  LSF (Load Sharing Facility) is a (licensed) product from Platform Computing Fairly distribute compute nodes among users enforce usage policies for established queues  most common queues: int, now, week, month RC uses Fair Share scheduling, not first come, first served (FCFS)  LSF commands typically start with the letter b (as in batch), e.g. bsub, bqueues, bjobs, bhosts, … see man pages for much more info!

41. 41 Simplified view of LSF bsub –n 64 –a mvapich –q week mpirun myjob Login Node Jobs Queued job routed to queue job_J job_F myjob job_7 job dispatched to run on available host which satisfies job requirements user logged in to login node submits job

42. 42 Running Programs on Topsail  Upon ssh to Topsail/Kure, you are on the Login node.  Programs SHOULD NOT be run on Login node.  Submit programs to one of the many, many compute nodes.  Submit jobs using Load Sharing Facility (LSF) via the bsub command.

43. 43 Common batch commands  bsub - submit jobs  bqueues – view info on defined queues bqueues –l week  bkill – stop/cancel submitted job  bjobs – view submitted jobs bjobs –u all  bhist – job history bhist –l

44. 44 Common batch commands  bhosts – status and resources of hosts (nodes)  bpeek – display output of running job  Use man pages to get much more info! man bjobs

45. 45 Submitting Jobs: bsub Command  Submit Jobs - bsub Run large jobs out of scratch space, smaller jobs can run out of your home space  bsub [-bsub_opts] executable [-exec_opts]  Common bsub options: –o  –o out.%J -q  -q week -R “resource specification”  -R “span[ptile=8]” -n  used for parallel, MPI jobs -a  -a mvapich(used on MPI jobs)

46. 46 Two methods to submit jobs:  bsub example: submit the executable job, myexe, to the week queue and redirect output to the file out. (default is to mail output)  Method 1: Command Line bsub –q week –o out.%J myexe  Method 2: Create a file (details to follow) called, for example, myexe.bsub, and then submit that file. Note the redirect symbol, < bsub < myexe.bsub

47. 47 Method 2 cont.  The file you submitted will contain all the bsub options you want in it, so for this example myexe.bsub will look like this #BSUB –q week #BSUB –o out.%J myexe  This is actually a shell script so the top line could be the normal #!/bin/csh, etc and you can run any commands you would like. if this doesn’t mean anything to you then nevermind :)

48. 48 Parallel Job example Batch Command Line Method  bsub –q week –o out.%J -n 64 -a mvapich mpirun myParallelExe Batch File Method  bsub < myexe.bsub  where myexe.bsub will look like this #BSUB –q week #BSUB –o out.%J #BSUB –a mvapich #BSUB –n 64 mpirun myParallelExe

49. 49 Some Topsail Queues QueueTime LimitJobs/UserCPU/Job int2 hrs128--- debug2 hrs64--- day24 hrs5124 – 128 week1 week5124 – 128 512cpu4 days51232 – 512 128cpu4 days51232 – 128 32cpu2 days5124 – 32 chunk4 days512Batch Jobs For access to the 512cpu queue the scalability must be demonstrated

50. 50 Some Kure Queues QueueTime LimitJobs/User int10 hrs2 debug5 minutes32 bigmem1 week8 week1 week- patronsnone- Most users have a 32 job slots limit unless they have been granted extra slots. Queues are always subject to change and probably will change as Kure production ramps up. Use the bqueues command to find the current status

51. 51 Common Error 1  If job immediately dies, check err.%J file  err.%J file has error: Can't read MPIRUN_HOST  Problem: MPI enivronment settings were not correctly applied on compute node  Solution: Include mpirun in bsub command

52. 52 Common Error 2  Job immediately dies after submission  err.%J file is blank  Problem: ssh passwords and keys were not correctly setup at initial login to Topsail  Solution: cd ~/.ssh/ mv id_rsa id_rsa-orig mv id_rsa.pub id_rsa.pub-orig Logout of Topsail Login to Topsail and accept all defaults

53. 53 Interactive Jobs  To run long shell scripts on Topsail or Kure, use int queue  bsub –q int –Ip /bin/bash This bsub command provides a prompt on compute node Can run program or shell script interactively from compute node

54. 54 Specialty Scripts  There are specialty scripts provided on Kure for the user convenience.  Batch scripts bmatlab, bsas, bstata  X-window scripts xmatlab, xsas, xstata  Interactive scripts imatlab, istata

55. 55 MPI/OpenMP Training  Courses are taught throughout year by Research Computing http://learnit.unc.edu/workshops http://help.unc.edu/CCM3_008194  See schedule for next course MPI OpenMP

56. 56 Further Help with Topsail/Kure  More details can be found on the Getting Started help documents: http://help.unc.edu/?id=6214 - Topsail http://help.unc.edu/?id=6214 http://help.unc.edu/ccm3_015682 - Kure http://help.unc.edu/ccm3_015682 http://keel.isis.unc.edu/wordpress/ - ON CAMPUS http://keel.isis.unc.edu/wordpress/  For assistance with Topsail/Kure, please contact the ITS Research Computing group Email: research@unc.eduresearch@unc.edu Phone: 919-962-HELP Submit help ticket at http://help.unc.eduhttp://help.unc.edu  For immediate assistance, see manual pages man