TinkerCliffs, ARC’s Flagship Resource

Overview

TinkerCliffs came online in the summer of 2020. With nearly 42,000 cores and over 93 TB of RAM, TinkerCliffs is nearly seven times the size of BlueRidge, ARC’s previous flagship CPU compute system, which was retired at the end of 2019. TinkerCliffs hardware is summarized in the table below.

Base Compute Nodes

DGX Nodes

A100 GPU Nodes

Intel Nodes

High Memory Nodes

Total

Vendor

Cray

Nvidia DGX A100

HPE Apollo 6500

Intel

Cray

-

Chip

AMD EPYC 7702

AMD EPYC 7742

AMD EPYC 7742

Intel Xeon Platinum 9242

AMD EPYC 7702

-

Nodes

308

10

4

16

8

346

Accelerators

-

8x NVIDIA A100-80G

8x NVIDIA A100-80G

-

-

-

Cores/Node

128

128

128

96

128

-

Memory (GB)/Node

256

2048

2048

384

1,024

-

Total Cores

39,424

1280

512

1,536

1,024

43,776

Total Memory (GB)

78,848

20,480

8,192

6,144

8,192

121,856

Local Disk

480GB SSD

30TB Gen4 NVMe

11.7TB NVMe

3.2TB NVMe

480GB SSD

-

Interconnect

HDR-100 IB

8x HDR-200 IB

4x HDR-200 IB

HDR-100 IB

HDR-100 IB

-

Tinkercliffs is hosted in the Steger Hall HPC datacenter on the Virginia Tech campus, so it is physically separated from other ARC HPC systems which are hosted in the AISB Datacenter at the Corporate Research Center (CRC) in Blacksburg.

For HPC, it is important that file systems (data storage) be physically near to the compute systems, so there is not direct connectivity from Tinkercliffs to some of the legacy filesystems (eg. GPFS /groups and /work). The /home filesystem on Tinkercliffs is the same as on legacy clusters, but for the reasons stated above, should not be used for i/o intensive workloads.

A BeeGFS file system was origally provided with the system to support /projects and /work filesystems for group collaboration and high-performance input/output (I/O), but was replaced in March 2022 with an IBM ESS GPFS filesystem supporting /projects. The /work is no longer available on Tinkercliffs with /fastscratch serving this role instead.

DGX Nodes - Added Fall 2022

Ten additional dense GPU nodes were acquired in 2022 to extend the capabilities of the Tinkercliffs cluster and made available to everyone in October 2022. These Nvidia DGX A100 nodes have very similar hardware specs to the four 2021 HPE Apollo 6500 “A100” nodes which were add to Tinkercliffs.

The DGX nodes are distinguished by their

  • dense, high-performance interconnecting network

  • the scale of resources which will permit multi-node jobs and more jobs per-user

  • larger and faster local scratch Gen4 NVMe array

  • a newer OS (RedHat Enterprise Linux 8.5) and linux kernel version 4.18.0-348.20.1.el8_5.x86_64 which will provide compatability with the latest software.

A100 GPU Nodes - Added Summer 2021

Four nodes nodes equipped with GPU accelerators were added to Tinkercliffs in June 2021. Each of these nodes is designed to be a clone of NVIDIA’s DGX nodes to provide a dense GPU resource for the VT research computing community. The eight NVIDIA A100-80G GPUs in each node are interconnected with NVIDIA’s NVLink technology. For internode communications, each chassis is equipped with four Mellanox HDR-200 Infiniband cards distributed across the PCIe Gen4 bus to provide each GPU with a nearby, high speed, low latency, path to the Infiniband network.

These nodes run CentOS 7.9 with Linux kernel version (3.10.0-1062.18.1.el7.x86_64).

Get Started

Tinkercliffs can be accessed via one of the two login nodes:

tinkercliffs1.arc.vt.edu tinkercliffs2.arc.vt.edu

For testing purposes, all users will be alloted 240 core-hours each month in the “personal” allocation. Researchers at the PI level are able to request resource allocations in the “free” tier (usage fully subsidized by VT) and can allocate 800,000 monthly billing units (normal_q core-hours) among their projects.

To do this, log in to the ARC allocation portal https://coldfront.arc.vt.edu,

  • select or create a project

  • click the “+ Request Resource Allocation” button

  • Choose the “Compute (Free) (Cluster)” allocation type

Usage needs in excess of 800,000 monthly billing units can be purchased via the ARC Cost Center.

Policies

Limits are set on the scale and quantity of jobs at the user and allocation (Slurm account) levels to help ensure availability of resources to a broad set of researchers and applications. These are the limits applied to free tier usage (note that the terms “cpu” and “core” are used interchangably here following Slurm terminology):

Policies for Main Usage Queues/Partitions

The normal_q, a100_normal_q, and dgx_normal_q are the partitions (queues) that handle the bulk of utilization on the Tinkercliffs cluster.

normal_q

a100_normal_q

dgx_normal_q

Node Type

Base Compute

A100 GPU

A100 GPU

Billing Weight

1

155.26/GPU

155.26/GPU

Number of Nodes

302

4

10

MaxRunningJobs (User)

24

12

32

MaxSubmitJobs (User)

240

48

64

MaxRunningJobs (Allocation)

48

24

48

MaxSubmitJobs (Allocation)

480

48

96

MaxNodes (User)

64

1

4

MaxNodes (Allocation)

96

2

6

MaxCPUs (User)

8192

128

512

MaxCPUs (Allocation)

12288

256

768

MaxGPUs (User)

-

8

32

MaxGPUs (Allocation)

-

16

48

MaxWallTime

6 days

6 days

6 days

Free allowance at Max[CPU/GPU]s (User)

4.07 days

26.84 days

6.71 days

Free allowance at Max[CPU/GPU]s (Alloc)

2.71 days

17.89 days

4.47 days

Policies for Development Usage Queues/Partitions

The “dev” partitions (queues) overlap the main usage queues above, but jobs in these queues get higher priority to allow more rapid access to resources for testing and development workloads. The tradeoff is that individuals may only run a small number of short jobs in these partitions.

dev_q

a100_dev_q

dgx_dev_q

Node Type

Base Compute

A100 GPU

A100 GPU

Billing Weight

1

155.26/GPU

155.26/GPU

Number of Nodes

307

4

10

MaxRunningJobs (User)

2

2

2

MaxSubmitJobs (User)

3

4

4

MaxRunningJobs (Allocation)

3

3

3

MaxSubmitJobs (Allocation)

6

6

6

MaxNodes (User)

64

1

4

MaxNodes (Allocation)

96

2

6

MaxCPUs (User)

8192

128

512

MaxCPUs (Allocation)

12288

256

768

MaxGPUs (User)

-

8

32

MaxGPUs (Allocation)

-

16

48

MaxWallTime

4 hours

4 hours

Free allowance at Max[CPU/GPU]s (User)

4.07 days

-

Free allowance at Max[CPU/GPU]s (Alloc)

2.71 days

-

Policies for Alternative Usage Queues/Partitions

The other four partitions on Tinkercliffs provide some alternative resources which can be useful

  • if you need a special type of resource such as larger memory or an Intel microarchitecture

  • when when the main partitions are very busy

  • reduced-rate or zero-billing queues if you have limited billing units in your account

preemptable_q

interactive_q

largemem_q

intel_q

Node Type

Base Compute

Base Compute

High Memory

Intel

Billing Weight

0 (no billing)

0.25/core

4.045454/core

3.772727/core

Number of Nodes

307

2

8

16

MaxRunningJobs (User)

64

4

4

8

MaxSubmitJobs (User)

640

4

40

80

MaxRunningJobs (Allocation)

128

-

6

12

MaxSubmitJobs (Allocation)

1280

-

60

120

MaxNodes (User)

-

-

4

8

MaxNodes (Allocation)

-

-

6

12

MaxCPUs (User)

-

64

512

768

MaxCPUs (Allocation)

-

128

768

1152

MaxGPUs (User)

-

-

-

-

MaxGPUs (Allocation)

-

-

-

-

MaxWallTime

-

4 hours

6 days

6 days

Free allowance at Max[CPU/GPU]s (User)

-

-

16.09 days

11.50 days

Free allowance at Max[CPU/GPU]s (Alloc)

-

-

10.73 days

7.67 days

Tinkercliffs is part of the ARC cost center, which provides a substantial “free tier” of usage. Each researcher is provided 800,000 billing units (1 billing unit = 1 TC normal_q core-hour) which can be divided among all projects and allocations they own. Monthly billing is based on usage attributed to jobs which complete in that month, so jobs which start in month A and finish in month B are billed in month B.

Modules

TinkerCliffs is different from previous ARC clusters in that it uses a new application stack/module system based on EasyBuild. Our old application stack was home-grown and involved a fair amount of overhead in getting new modules - e.g., new versions of a package - installed. EasyBuild streamlines a lot of that work and should also make it trivial in some cases for users to install their own versions of packages if they so desire. Key differences from a user perspective include:

  • Hierarchies are replaced by toolchains. Right now, there are two:

    • foss (“Free Open Source Software”): gcc compilers, OpenBLAS for linear algebra, OpenMPI for MPI, etc

    • intel: Intel compilers, Intel MKL for linear algebra, Intel MPI

  • Instead of loading modules individually (e.g., module load intel mkl impi), a user can just load the toolchain (e.g., module load intel/2019b).

  • Modules load their dependencies, e.g.,

$ module reset; module load HPL/2.3-intel-2019b; module list 
  
  Currently Loaded Modules:
    1) shared                              6) craype-x86-rome            11) binutils/2.32-GCCcore-8.3.0          16) intel/2019b
    2) slurm/20.02.3                       7) craype-network-infiniband  12) iccifort/2019.5.281                  17) HPL/2.3-intel-2019b
    3) apps                                8) DefaultModules             13) impi/2018.5.288-iccifort-2019.5.281
    4) site/tinkercliffs/easybuild/setup   9) GCCcore/8.3.0              14) iimpi/2019b
    5) cray                               10) zlib/1.2.11-GCCcore-8.3.0  15) imkl/2019.5.281-iimpi-2019b
  • All modules are visible with module avail. So in many cases it is probably better to search with module spider rather than printing the whole list.

  • Some key system software, like the Slurm scheduler, are included in default modules. This means that module purge can break important functionality. Use module reset instead.

  • Lower-level software is included in the module structure (see, e.g., binutils in the HPL example above), which should mean less risk of conflicts in adding new versions later.

  • Environment variables (e.g., $SOFTWARE_LIB) available in our previous module system may not be provided. Instead, EasyBuild typically provides $EBROOTSOFTWARE to point to the software installation location. So for example, to link to NetCDF libraries, one might use -L$EBROOTNETCDF/lib64 instead of the previous -L$NETCDF_LIB.

Architecture

  • The AMD Rome architecture is similar to Cascades in that it is x86_64 but lacks the AVX-512 instruction set added to Intel processors in the last couple of years.

  • Nodes are larger (128 cores) and have more memory bandwidth (~350 GB/s).

  • There are eight NUMA (memory locality) domains per node and one L3 cache for every four cores.

Optimization

See also the tuning guides available at https://developer.amd.com/, especially this guide to compiler flags.

  • Cache locality really matters - process pinning can make a big difference on performance.

  • Hybrid programming often pays off - one MPI process per L3 cache with 4 threads is often optimal.

Intel toolchain:

  • Fast, though our testing has found that v2020 is slower than v2019

  • Avoid –xhost

  • Use -march=core-avx2 to get the optimal vectorization instruction set

  • Use the following environment variables for MKL (we set these as part of the MKL module):

  export MKL_DEBUG_CPU_TYPE=5
  export MKL_ENABLE_INSTRUCTIONS=AVX2

Foss (GCC) toolchain:

  • Use -mtune=znver2 -march=znver2 to target the Zen2 architecture

  • Use -mavx2 to get the optimal vectorization instruction set

AOCC Compiler:

  • AMD compiler. Very fast on Rome architectures. ARC is working on getting AOCC integrated into a toolchain.

  • Use -mtune=znver2 -march=znver2 to target the Zen2 architecture

  • Use -mavx2 to get the optimal vectorization instruction set

Examples

See below for a series of examples of how to compile code for a variety of compilers and for how to run optimally in a variety of configurations. These and a wide variety of simple application-specific examples can be found in our examples repository.

Stream

STREAM is a memory bandwidth benchmark. To maximize bandwidth, we run in parallel with one process per L3 cache (cores 0, 4, …, 124).

  #Load the Intel toolchain
  module reset; module load intel/2019b
  
  #Tell OpenMP to use every 4th core
  export OMP_PROC_BIND=true
  export OMP_NUM_THREADS=32
  export OMP_PLACES="$( seq -s },{ 0 4 127 | sed -e 's/\(.*\)/\{\1\}/' )"
  
  #Compile
  icc -o stream.intel stream.c -DSTATIC -DNTIMES=10 -DSTREAM_ARRAY_SIZE=2500000000 \
    -mcmodel=large -shared-intel -Ofast -qopenmp -ffreestanding -qopt-streaming-stores always
  
  #Run
  ./stream.intel

Results:

  Function    Best Rate MB/s
  Copy:             341475.1 
  Scale:            341770.0 
  Add:              336668.3 
  Triad:q:          336972.6 

MT-DGEMM

mt-dgemm is a threaded matrix multiplication program that can be used to benchmark dense linear algebra libraries. Here we use it to show how to link against linear algebra libraries and run efficiently across a socket.

AOCC

#Load the aocc and blis modules
module reset; module load aocc/aocc-compiler-2.1.0 amd-blis/aocc/64/2.1

#Compile:
#  Build for the Rome architecture: -mtune=znver2 -march=znver2
#  Use fast vectorization: -mavx2
#  Use math libraries: -lm
#  Use OpenMP: -fopenmp -lomp
#  Other optimizations: -Ofast -ffp-contract=fast -funroll-loops
#  Link with AMD BLIS linear algebra library: -I$BLISDIR/../include $BLISDIR/libblis-mt.a
#  Macro used by the mt-dgemm program: -D USE_CBLAS
clang -mtune=znver2 -march=znver2 -mavx2 -lm -fopenmp -lomp -Ofast -ffp-contract=fast -funroll-loops -I$BLISDIR/../include $BLISDIR/libblis-mt.a -D USE_CBLAS -o mt-dgemm.aocc mt-dgemm.c

#Run with 64 OpenMP threads on cores 0-63 (socket 1) using NUMA memory regions 0-3 (socket 1). This keeps Linux from moving the threads away from memory.
OMP_NUM_THREADS=64 GOMP_CPU_AFFINITY=0-63:1 numactl --membind=0-3 ./mt-dgemm.aocc 16000

GCC

#Load the foss toolchain
module reset; module load foss/2020a

#Compile:
#  Build for the Rome architecture: -mtune=znver2 -march=znver2
#  Use fast vectorization: -mavx2
#  Use math libraries: -lm
#  Use OpenMP: -fopenmp
#  Other optimizations: -Ofast -ffp-contract=fast -funroll-loops
#  Link with OpenBLAS linear algebra library: -L$OPENBLAS_LIB -lopenblas
#  Macro used by the mt-dgemm program: -D USE_CBLAS
gcc -mtune=znver2 -march=znver2 -mavx2 -lm -fopenmp -Ofast -ffp-contract=fast -funroll-loops -L$OPENBLAS_LIB -lopenblas -D USE_CBLAS -o mt-dgemm.gcc mt-dgemm.c

#Run with 64 OpenMP threads on the cores (0-63) and memory (regions 0-3) associated with socket 1. This keeps Linux from moving the threads away from memory. Using GOMP_CPU_AFFINITY to pin thread 0 to core 0, thread 1 to core 1, etc would be ideal but breaks the threading in OpenBLAS for whatever reason.
OMP_NUM_THREADS=64 numactl -C 0-63 --membind=0-3 ./mt-dgemm.gcc 16000

Intel

Here we use intel 2019 as testing indicates that 2020 is substantially slower.

#Load the intel toolchain
module reset; module load intel/2019b

#Note that the module has set MKL_ENABLE_INSTRUCTIONS=AVX2 and MKL_DEBUG_CPU_TYPE=5
 to ensure that MKL uses the optimal instruction set
env | egrep "MKL_DEBUG_CPU_TYPE|MKL_ENABLE_INSTRUCTIONS"

#Compile:
#  Use fast vectorization: -march=core-avx2
#  Use OpenMP: -qopenmp
#  Other optimizations: -O3 -ffreestanding
#  Link with MKL linear algebra library: -mkl
#  Macro used by the mt-dgemm program: -D USE_MKL=1
icpc -march=core-avx2 -qopenmp -O3 -ffreestanding -mkl -D USE_MKL=1 -o mt-dgemm.intel mt-dgemm.c 

#Run with 64 threads on cores 0-63 (socket 1) using NUMA memory regions 0-3 (socket 1). This keeps Linux from moving the threads away from memory.
MKL_NUM_THREADS=64 GOMP_CPU_AFFINITY=0-63:1 numactl --membind=0-3 ./mt-dgemm.intel 16000

Results

The results show the benefits of AMD’s optimizations and of MKL’s performance over OpenBLAS:

  aocc+blis 2.1:  1658.861832 GF/s
  foss/2020a:     1345.527671 GF/s
  intel/2019b:    1615.846327 GF/s

HPL

HPL is a computing benchmark. Here we use it to demonstrate how to run in the pure MPI (1 process per core) and hybrid MPI+OpenMP (1 process per L3 cache with 4 OpenMP threads working across the cache) models. To load the HPL module, we can do simply

  module reset; module load HPL/2.3-intel-2019b  #intel
  module reset; module load HPL/2.3-foss-2020a   #gcc

MPI Only (1 MPI process/core)

Here we use pure MPI and start one MPI process per core. Jobs in this case should typically be requested with –ntasks-per-node=128 (if you want full node performance).

  • Intel, using mpirun. We use an environment variable to make sure that MPI processes are laid out in order and not moved around by the operating system.

  mpirun -genv I_MPI_PIN_PROCESSOR_LIST=0-127 xhpl
  • gcc, using mpirun. Here we use OpenMPI’s mapping and binding functionality to assign the processes to consecutive cores.

  mpirun --map-by core --bind-to core -x OMP_NUM_THREADS=1 xhpl
  • Intel or gcc, using srun. We use srun’s cpu-bind flag to bind the processes to cores.

  srun --cpu-bind=cores xhpl

Hybrid MPI+OpenMP (1 MPI process/L3 cache)

Here we start one MPI process per L3 cache (every 4 cores). Jobs in this case should typically be requested with –ntasks-per-node=32 –cpus-per-task=4 so that Slurm knows how many processes you need.

  • Intel, using mpirun. We use environment variables to tell mpirun to start a process on every fourth core and use 4 OpenMP (MKL) threads per process:

  mpirun -genv I_MPI_PIN_PROCESSOR_LIST="$( seq -s , 0 4 127 )" -genv I_MPI_PIN_DOMAIN=omp -genv OMP_NUM_THREADS=4 -genv OMP_PROC_BIND=TRUE -genv OMP_PLACES=cores xhpl
  • gcc, using mpirun. Here we use OpenMPI’s mapping and binding functionality to assign the processes to L3 caches.

  mpirun --map-by ppr:1:L3cache --bind-to l3cache -x OMP_NUM_THREADS=4 xhpl
  • Intel or gcc, using Slurm’s srun launcher. We use a cpu mask to tell Slurm which cores each process should have access to. (0xF is hexadecimal for 15, or 1111 in binary, meaning access should be allowed to the first four cores. 0xF0 is 11110000 in binary, meaning access should be allowed to the second set of four cores. The list continues through 11110000…..0000, indicating that the last process should have access to cores 124-127.)

  srun --cpu-bind=mask_cpu=0xF,0xF0,0xF00,0xF000,0xF0000,0xF00000,0xF000000,0xF0000000,0xF00000000,0xF000000000,0xF0000000000,0xF00000000000,0xF000000000000,0xF0000000000000,0xF00000000000000,0xF000000000000000,0xF0000000000000000,0xF00000000000000000,0xF000000000000000000,0xF0000000000000000000,0xF00000000000000000000,0xF000000000000000000000,0xF0000000000000000000000,0xF00000000000000000000000,0xF000000000000000000000000,0xF0000000000000000000000000,0xF00000000000000000000000000,0xF000000000000000000000000000,0xF0000000000000000000000000000,0xF00000000000000000000000000000,0xF000000000000000000000000000000,0xF0000000000000000000000000000000 xhpl

Results

The results show the benefit of the hybrid MPI+OpenMP model and of MKL over OpenBLAS, particularly in the hybrid model.

  intel |     mpi  | mpirun | 2,944 GFlops/s
  intel |     mpi  |   srun | 2,809 GFlops/s
    gcc |     mpi  | mpirun | 2,734 GFlops/s
    gcc |     mpi  |   srun | 2,659 GFlops/s
  intel | mpi+omp  | mpirun | 3,241 GFlops/s
  intel | mpi+omp  |   srun | 3,227 GFlops/s
    gcc | mpi+omp  | mpirun | 2,836 GFlops/s
    gcc | mpi+omp  |   srun | 2,845 GFlops/s