UG3.3: GALILEO100 UserGuide

System Architecture

Compute Nodes:

• 554 computing nodes each 2 x CPU Intel CascadeLake 8260,  with 24 cores each, 2.4 GHz, 384GB RAM, subdivided in:
  
  • • 340 standard nodes ("thin nodes") 480 GB SSD
    • 180  data processing nodes ("fat nodes") 2TB SSD, 3TB Intel Optain
    •   34 (visualization "viz" ) GPU nodes with 2x NVIDIA GPU V100 with 100Gbs Infiniband interconnection and 2TB SSD.
• 77 computing server OpenStack for cloud computing, 2x CPU 8260 Intel CascadeLake, 24 cores, 2.4 GHz, 768 GB RAM, with 100Gbs Ethernet interconnection.
• 20 PB of active storage accessible from both cloud and HPC nodes.
• 5 PB of fast storage for HPC system.
• 1 PB Ceph storage for Cloud (full NVMe/SSD)
• 720 TB fast storage (IME DDN solution)

Login and Service nodes:

10 login nodes and 5 service nodes. All the nodes are interconnected through an Infiniband network, with OPA v10.6, capable of a maximum bandwidth of 100Gbit/s between each pair of nodes.

Accounting

For more information about accounting, please consult our dedicated section.

Budget Linearization policy

On GALILEO100 a linearization policy for the usage of project budgets has been defined and implemented. For each account, a monthly quota is defined as:

monthTotal = (total_budget / total_no_of_months)

Starting from the first day of each month, the collaborators of any account are allowed to use the quota at full priority. As long as the budget is consumed, the jobs submitted from the account will gradually lose priority, until the monthly budget (monthTotal) is fully consumed. At that moment, their jobs will still be considered for execution, but with a lower priority than the jobs from accounts that still have some monthly quota left.

This policy is similar to those already applied by other important HPC centers in Europe and worldwide. The goal is to improve the response time, giving users the opportunity of using the cpu hours assigned to their project in relation to their actual size (total amount of core-hours).

Disks and Filesystems

The storage organisation conforms to the CINECA infrastructure (see Section "Data storage and Filesystems") . In addition to the home directory ($HOME), for each user is defined a scratch area $CINECA_SCRATCH, a large disk for storing run time data and files. $WORK is defined for each active project on the system, reserved for all the collaborators of the project. This is a safe storage area to keep run time data for the whole life of the project.

The filesystem organisation is based on LUSTRE an open source parallel file system.

It is also available a temporary storage local on compute nodes generated when the job starts and accessible via environment variable $TMPDIR. For more details please see the dedicated section of UG2.5: Data storage and FileSystems. On Galileo100 the $TMPDIR local area has 293 GB of available space.

$DRES points to the shared repository where Data RESources are maintained. This is a data archive area available only on-request, shared with all CINECA HPC systems and among different projects.

$DRES is not mounted on the compute nodes. This means that you can't access it within a batch job: all data needed during the batch execution has to be moved on $WORK or $CINECA_SCRATCH before the run starts.

Use the local command "cindata" to query for disk usage and quota ("cindata -h" for help):

  > cindata

Modules environment

The software modules are collected in different profiles and organized by functional category (compilers, libraries, tools, applications,..).

On GALILEO100 the profiles are of two types, “domain” type (bioinf, chem-phys, lifesc,..) for the production activity and “programming” type (base and advanced) for compilation, debugging and profiling activities and that they can be loaded together.

"Base" profile is the default. It is automatically loaded after login and it contains basic modules for the programming activities (intel e gnu compilers, math libraries, profiling and debugging tools,..).

If you want to use a module placed under other profiles, for example an application module, you will have to load preventively the corresponding profile:

>module load profile/<profile name>

>module load autoload <module name>

For listing all profiles you have loaded you can use the following command:

>module list

In order to detect all profiles, categories and modules available on GALILEO100 the command “modmap” is available:

>modmap

With modmap you can see if the desired module is available and which profile you have to load to use it.

>modmap -m <module name>

Spack environment - will be available soon

In case you don't find a software you are interested in, you can install it by yourself. 
In this case, on GALILEO100 we also offer the possibility to use the “spack” environment by loading the corresponding module. Please refer to the dedicated section in UG2.6: Production Environment.

Production environment

Since GALILEO100 is a general purpose system and it is used by several users at the same time, long production jobs must be submitted using a queuing system. This guarantees that access to the resources is as fair as possible.

Roughly speaking, there are two different modes to use an HPC system: Interactive and Batch. For a general discussion see the section "Production Environment".

Interactive

A serial program can be executed in the standard UNIX way:

> ./program

This is allowed only for very short runs, since the interactive environment set on the login nodes has a 10 minutes time limit: for longer runs please use the "batch" mode.

A parallel program can be executed interactively only by submitting an "Interactive" SLURM batch job, using the "srun" command: the job is queued and scheduled as any other job, but when executed, the standard input, output, and error streams are connected to the terminal session from which srun was launched.

For example, to start an interactive session with the MPI program "myprogram", using one node and two processors, you can launch the command:

> salloc -N 1 --ntasks-per-node=2 -A <account_name> 

SLURM will then schedule your job to start, and your shell will be unresponsive until free resources are allocated for you. If not specified, the default time limit for this kind of jobs is one hour.

When the shell returns a prompt inside the compute node, you can execute your program by typing:

> srun ./myprogram

(srun is recommended with respect to mpirun for this environment)

SLURM automatically exports the environment variables you defined in the source shell, so that if you need to run your program "myprogram" in a controlled environment (i.e. with specific library paths or options), you can prepare the environment in the login shell and be sure to find it again in the interactive shell o the compute node.

On systems using SLURM, you can submit a script script.x using the command:

> sbatch script.x

You can get a list of defined partitions with the command:

> sinfo

For more information and examples of job scripts, see section Batch Scheduler SLURM.

Submitting serial Batch Jobs

The  partition will be  available in the full production.

Graphic session

If a graphic session is desired we recommend to use the tool "RCM". Please install the latest version of RCM. See the corresponding paragraph to know more about how to download and use RCM.

Submitting parallel Batch Jobs

To run parallel batch jobs on GALILEO100 you need to specify the partition  and the qos that are described in this user guide.

If you do not specify the partition, your jobs will try to run on the default partition  g100_usr_prod.

The minimum number of cores you can request for a batch job is 1. The maximum number of cores that you can request is  (16 nodes). It is also possible to request a maximum walltime of 24 hours. Defaults are as follows:

• If you do not specify the walltime (by means of the #SBATCH --time directive), a default value of 30 minutes will be assumed.

• If you do not specify the number of cores (by means of the "SBATCH -n" directive) a default value of 1 core will be assumed.

• If you do not specify the amount of memory (as the value of the "SBATCH --mem" DIRECTIVE), a default value of 7800 MBper core will be assumed.

The maximum memory per node is 375300MB (366.5GB) for thin and viz nodes, about 3TB for fat nodes.

Processor affinity

Processor affinity, or CPU pinning, enables the binding of processes and threads to a CPU (or group of CPUs). It is crucial to ensure the correct affinity so to avoid the CPUs overallocation, with a significant reduction of performances. It becomes a critical matter when you ask for a full node but, for your specific reasons (memory needs etc.) you don't use all the cores.  

The following indications work when running your executables with srun, which is the recommended option against mpirun. We refer to a hybrid MPI/OpenMP case compiled with the Intel oneAPI suite.

Given your optimal value of OMP_NUM_THREADS and number of processes, to obtain the full node ask for a number of task such that  ( --ntasks-per-node * --cpus-per-task )= 48.

• To avoid the processes over allocation of cores rely on the --cpu-bind=cores option of srun  (you can skip it if you use all the requested cores)
• To enforce the threads affinity use the Intel parameter KMP_AFFINITY, or the OpenMP parameter OMP_PLACES
• To distribute the MPI tasks consecutively inside the sockets, use the -m block:block option of srun (or the equivalent sbatch directive #SBATCH -m block:block)

#!/bin/bash
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=12
#SBATCH --cpus-per-task=4
#SBATCH --account=<your_account>

module load autoload intelmpi/oneapi-2021–binary

export OMP_NUM_THREADS=4

export KMP_AFFINITY=compact    # or OMP_PLACES=cores

srun --cpu-bind=cores -m block:block <your_exe>

Use of GPUs on Galileo100

to be soon defined

Users with reserved resources

Users of projects that require reserved resources (such as industrial users or users associated to an agreement that involves dedicated resources) will be associated to a QOS qos_ind.

Using the  qos_ind (i.e. specifying the QOS in the submission script) , and specifying the partition g100_spc_prod, users associated to the allowed project will run their jobs on reserved nodes in the g100_spc_prod partition with the features and limits imposed for the particular account.

>#SBATCH --partition=g100_spc_prod

>#SBATCH --qos=qos_ind

Summary

In the following table, you can find all the main features and limits imposed on the SLURM partitions and QOS.

Intel Compilers

The native, and recommended, compilers on GALILEO100 are the Intel ones, since the architecture is based on Intel processors and therefore using the Intel compilers may result in a significant improvement in performance and stability of your code. On the cluster is installed the new suite Intel OneAPI. Initialize the environment with the module command:

> module load intel/oneapi-2021–binary

> module list
   Currently Loaded Modulefiles:
   intel/oneapi-2021–binary

The suite contains the new Intel oneAPI nextgen compilers (icx, icpx, ifx) and the classic compilers (icc, icpc, ifort):

NOTE:

• ICX is a new compiler. It has functional and behavioral differences compared to ICC. You can expect some porting will be needed for existing applications using ICC. According to Intel, the transition from ICC Classic to ICX is smooth and effortless. However, you must port and tune any existing applications from ICC Classic to ICX. Please refer to the official Intel Porting Guide for ICC Users to DPCPP or ICX
• IFORT is a completely new compiler. According to Intel, although considerable effort is being made to make the transition from ifort to ifx as smooth and as effortless as possible, customers can expect that some effort may be required to tune their application. IFORT will remain Intel’s recommended production compiler until ifx has performance and features superior to ifort. Please refer to the official Intel Porting Guide for ifort Users to ifx

The documentation can be obtained with the man command after loading the relevant module:

> man ifort
> man icx

Some miscellaneous flags are described in the following:

-extend_source    Extend over the 77 column F77's limit 
-free / -fixed    Free/Fixed form for Fortran
-ip               Enables additional interprocedural optimization for single-file compilation
-ipo              Enables interprocedural optimization between files - whole program optimisation

NOTE for the migration from Galileo to Galileo100: In principle, binaries generated on Galileo should work, but we strongly recommend you to reinstall all your software applications since on Galileo100 there is a different Operating System (Centos 8.3).

GNU compilers

The gnu compilers are always available but they are not the best optimizing compilers, especially for an IntelOneAPI-based cluster like GALILEO100.

For a more recent version of the compiler, initialize the environment with the module command:

> module load gnu

The name of the GNU compilers are:

• g77: Fortran77 compiler
• gfortran: Fortran compiler with "gnu" standard
• gcc: C compiler
• g++: C++ compiler

The "gnu" standard is the default value for the -std option. It specifies a superset of the latest Fortran standard that includes all of the extensions supported by GNU Fortran, although warnings will be given for obsolete extensions not recommended for use in new code. To change the standard to which the program is expected to conform, set the -std option to one of the possible values (f95, f2003, f2008, f2018, gnu, or legacy).

The documentation can be obtained with the man command:

> man gfortan
> man gcc

Some miscellaneous flags are described in the following:

-ffixed-line-length-132       To extend over the 77 column F77's limit
-ffree-form / -ffixed-form    Free/Fixed form for Fortran

Debuggers and Profilers

If your code aborts at runtime, there may be a problem with it. In order to solve it, you can decide to analyze the core file or to run your code using a debugger.

-O0     Lower level of optimisation
-g      Produce debugging information

-traceback        generate extra information to provide source file traceback at run time
-fp-stack-check   generate extra code to ensure that the floating-point stack is in the expected state
-check bounds     enables checking for array subscript expressions
-fpe0             allows some control over floating-point exception handling at run-time

-Wall             Enables warnings pertaining to usage that should be avoided
-fbounds-check    Checks for array subscripts.

> module load gnu
> gfortran -O0 -g -Wall -fbounds-check -o myexec myprog.f90
> gdb ./myexec

> ulimit -c unlimited (bash)
> limit coredumpsize unlimited (csh/tcsh)

> export decfort_dump_flag=TRUE  (bash)       
> setenv decfort_dump_flag TRUE  (csh/tcsh)

> module load intel
> ifort -O0 -g -traceback -fp-stack-check -check bounds -fpe0 -o myexec prog.f90
> ulimit -c unlimited
> export decfort_dump_flag=TRUE
> ./myexec
> ls -lrt
  -rwxr-xr-x 1 aer0 cineca-staff   9652 Apr  6 14:34 myexec
  -rw------- 1 aer0 cineca-staff 319488 Apr  6 14:35 core.25629
> idbc ./myexec core.25629

> module load gnu
> gfortran -O0 -g -Wall -fbounds-check -o myexec prog.f90
> ulimit -c unlimited
> ./myexec
> ls -lrt
  -rwxr-xr-x 1 aer0 cineca-staff   9652 Apr  6 14:34 myexec
  -rw------- 1 aer0 cineca-staff 319488 Apr  6 14:35 core.25555
> gdb ./myexec core.2555

Totalview - will be soon available
Totalview is a parallel debugger with a practical GUI that assist users to debug their parallel code. It has functionalities like stopping and reprising a code mid-run, setting breakpoints, checking the value of variables anytime, browse between the different tasks and threads to see the different behaviours, memory check functions and so on. For information about how to run the debugger (by connecting the compute nodes to your display via RCM), type the command:

> module help totalview

Scalasca
Scalasca is a tool for profiling parallel scientific and engineering applications that make use of MPI and OpenMP.
Details how to use scalasca in
http://www.scalasca.org/software/scalasca-2.x/documentation.html

Profilers - it will be updated

In software engineering, profiling is the investigation of a program's behaviour using information gathered as the program executes. The usual purpose of this analysis is to determine which sections of a program to optimize - to increase its overall speed, decrease its memory requirement or sometimes both.

A (code) profiler is a performance analysis tool that, most commonly, measures only the frequency and duration of function calls, but there are other specific types of profilers (e.g. memory profilers) in addition to more comprehensive profilers, capable of gathering extensive performance data.

Scientific libraries (MKL)

MKL

The Intel Math Kernel Library (Intel MKL) enables improving performance of scientific, engineering, and financial software that solves large computational problems. Intel MKL provides a set of linear algebra routines, fast Fourier transforms, as well as vectorized math and random number generation functions, all optimized for the latest Intel processors, including processors with multiple cores.

Intel MKL is thread-safe and extensively threaded using the OpenMP technology.

documentation can be found by loading the mkl module and searching in the directory:

${MKLROOT}/../Documentation/en_US/mkl

To use the MKL in your code you to load the module, then to define includes and libraries at compile and linking time:

> module load mkl
> icc -I$MKL_INC -L$MKL_LIB  -lmkl_intel_lp64 -lmkl_core -lmkl_sequential

For more informations please refer to the documentation.