|
|
| (3 intermediate revisions by 2 users not shown) |
| Line 1: |
Line 1: |
| [[File:Cluster_scheme.png|thumb|right|600px|Schematic overview of the cluster.]]
| | <!-- TODO: several specifics on this page need confirmation against the current cluster — node counts, memory sizes, exact interconnect, and a current architecture diagram. Search for "TODO" below. --> |
| The new Agrogenomics HPC has a classic cluster architecture: state of the art Parallel File System (PSF), headnodes, compute nodes (of varying 'size'), all connected by superfast network connections (Infiniband). Implementation of the cluster will be done in stages. The initial stage includes a 600TB PFS, 48 slim nodes of 16 cores and 64GB RAM each, and 2 fat nodes of 64 cores and 1TB RAM each. The overall architecture, that include two head nodes in fall-over configuration and an infiniband network backbone, can be easily expanded by adding nodes and expanding the PFS. The cluster management software is designed to facilitate a heterogenous and evolving cluster.
| |
| {{-}}
| |
| === Nodes ===
| |
| The cluster consists of a bunch of separate machines that each has its own operating system. The default operating system throughout the cluster is [https://www.scientificlinux.org Scientific Linux] version 6. Scientific Linux (SL) is based on [http://en.wikipedia.org/wiki/Red_Hat_Enterprise_Linux Red Hat Enterprise Linux (RHEL)], which currently is at version 6. SL therefore follows the versioning scheme of RHEL.
| |
|
| |
|
| The cluster has two master nodes in a redundant configuration, which means that if one crashes, the other will take over seamlessly. Various other nodes exist to support the two main file systems (the Lustre parallel file system and the NFS file system). The actual computations are done on the worker nodes or compute nodes. The cluster is configured in a heterogeneous fashion: it consists of 48 so called 'slim nodes', that each have 16 cores and 64GB of RAM (called 'node001' through 'node060'; note that not all node names map to physical nodes), and two so called 'fat nodes' that each have 64 cores and 1TB of RAM ('fat001' and 'fat002').
| | Anunna has a classic, heterogeneous HPC architecture: a set of login nodes that users connect to, a larger pool of compute nodes where the actual work runs, a fast parallel filesystem shared across the whole cluster, and a high-speed network tying it all together. The design is built to grow — nodes and storage can be added over time. |
|
| |
|
| Information from the Cluster Management Portal, as it appeared on June 26, 2014:
| | == Login nodes == |
| <code>DEVICE INFORMATION
| |
| Hostname State Memory Cores CPU Speed GPU NICs IB Category
| |
| node001..node002 UP 67.6 GiB 16 Intel(R) Xeon(R) CPU E5-2660 0+ 2200 MHz 3 1 default
| |
| node049..node054 UP 67.6 GiB 16 Intel(R) Xeon(R) CPU E5-2660 0+ 2200 MHz 3 1 default
| |
| master1 master2 UP 67.5 GiB 16 Intel(R) Xeon(R) CPU E5-2660 0+ 2199 MHz 5 1 custom
| |
| mds01, mds02 UP 16.8 GiB 8 Intel(R) Xeon(R) CPU E5-2609 0+ 2399 MHz 5 1 mds
| |
| storage01..storage06 UP 67.6 GiB 32 Intel(R) Xeon(R) CPU E5-2660 0+ 2200 MHz 5 1 oss
| |
| nfs01 UP 67.6 GiB 8 Intel(R) Xeon(R) CPU E5-2609 0+ 2399 MHz 7 1 login
| |
| fat001 fat002 UP 1.0 TiB 64 AMD Opteron(tm) Processor 6376 2300 MHz 5 1 fat
| |
| </code>
| |
|
| |
|
| Main cluster node configuration:
| | You reach the cluster by connecting to <code>login.anunna.wur.nl</code> over SSH (see [[Log in to Anunna]]). The login nodes are the entry point: you use them to edit files, manage your data, and submit jobs. They are shared by everyone, so they are not for heavy computation — anything beyond light, interactive work belongs in a job (see [[Batch Jobs]] and [[Interactive Jobs]]). |
| * Master nodes: 2 PowerEdge R720 master nodes in a failover configuration, which also will share some applications and databases with machines in the cluster, for which the parallel file system is not the ideal solution.
| |
| * The NFS server is a PowerEdge R720XD. The NFS node will also act as a login node, where users log in and compile applications and submit jobs and share each home directory via nfs.
| |
| * 50 compute nodes
| |
| ** 12x Dell PowerEdge C6000 enclosures, each containing four nodes
| |
| ** 48x Dell PowerEdge C6220; 16 Intel Xeon cores, 64GB RAM each
| |
| ** 2x Dell R815; 64 AMD Opteron cores, 1TB RAM each
| |
| Hyperthreading is disabled in compute nodes.
| |
|
| |
|
| === Filesystems === | | == Compute nodes == |
|
| |
|
| [[File:Storage_pic.png|thumb|right|300px|Schematic overview of storage components of the B4F cluster.]]
| | The compute nodes do the actual work. They are heterogeneous: ordinary CPU nodes for most jobs, plus nodes equipped with NVIDIA or AMD GPUs for accelerated workloads. Jobs reach them through the SLURM scheduler, which places each job on a node that matches its requested resources. |
| The Agrogenomics Cluster has two primary file systems, each with different properties and purposes. | |
| ==== Parallel File System: Lustre ====
| |
| At the base of the cluster is an ultrafast file system, a so called [http://en.wikipedia.org/wiki/Parallel_file_system Parallel File System] (PFS). The current size of the PFS is around 600TB. The PFS implemented in the Agrogenomics Cluster is called [http://en.wikipedia.org/wiki/Lustre_(file_system) Lustre]. Lustre has become very popular in recent years due to the fact that it is very feature rich, deemed very stable, and is Open Source. Lustre nowadays is the default option for PFS in Dell clusters as well as clusters sold by other vendors. The PFS is mounted on all head nodes and worker nodes of the cluster, providing a seamless integration between compute and data infrastructure. The strength of a PFS is speed - the total I/O should be up to 15GB/s. by design. With a very large number of compute nodes - and with very high volumes of data - these high read-write speeds that the PSF can provide are necessary. The Lustre filesystem is divided in [[Lustre_PFS_layout | several partitions]], each differing in persistence and backup features. The Lustre PSF is meant to store (shared) data that is likely to be used for analysis in the near future. Personal analysis scripts, software, or additional small data files can be stored in the $HOME directory of each of the users.
| |
|
| |
|
| The hardware components of the PFS:
| | * CPU work runs in the <code>main</code> partition. |
| * 2x Dell PowerEdge R720 | | * GPU work runs in the <code>gpu</code> (NVIDIA) and <code>gpu_amd</code> (AMD) partitions. |
| * 1x Dell PowerVault MD3220
| |
| * 6x Dell PowerEdge R620
| |
| * 6x Dell PowerVault MD3260
| |
|
| |
|
| ==== Network File System (NFS): $HOME dirs ====
| | For the hardware details of each node type, see [[Compute Hardware Overview]]; for how to select a partition or constrain a job to particular hardware, see [[Partitions / Queues]] and [[Choosing a node (constraints)]]. |
| Each user will have his/her own home directory. The path of the home directory will be:
| |
|
| |
|
| /home/[name partner]/[username]
| | == Storage == |
|
| |
|
| /home lives on a so called [http://en.wikipedia.org/wiki/Network_File_System Network File System], or NFS. The NFS is separate from the PFS and is far more limited in I/O (read/write speeds, latency, etc) than the PFS. This means that it is not meant to store large datavolumes that require high data transfer or small latency. Compared to the Lustre PFS (600TB in size), the size of the NFS is small in comparison - only 20TB. The /home partition will be backed up daily. The amount of space that can be allocated is limited per user. Personal quota and total use per user can be found using (200GB soft and 210GB hard limit) :
| | A [https://en.wikipedia.org/wiki/Lustre_(file_system) Lustre] parallel filesystem (mounted at <code>/lustre</code>) is shared across all nodes and is the cluster's main working storage; your home directory lives on a separate, smaller NFS filesystem. Longer-term and archival storage is available too. For the full picture — quotas, backup behaviour, and which tier to use for what — see [[Storage Systems Overview]]. |
|
| |
|
| <source lang='bash'>
| | == Network == |
| quota -s
| |
| </source>
| |
|
| |
|
| The NFS is supported through the NFS server (nfs01) that also serves as access point to the cluster. | | The nodes and the Lustre filesystem are connected by a high-speed, low-latency interconnect, which is what allows the parallel filesystem and multi-node jobs to perform well. <!-- TODO: confirm current interconnect (Omnipath and/or InfiniBand) and topology. --> |
|
| |
|
| Hardware components of the NFS:
| | == Operating system == |
| * 1x Dell PowerEdge R720XD
| |
| * 1x Dell PowerVault MD3220
| |
|
| |
|
| === Network === | | The cluster runs Ubuntu 24.04 across its nodes. Software is provided through [[Environment Modules | environment modules]] rather than being installed system-wide, so several versions of the same program can coexist. |
| The various components - head-nodes, worker nodes, and most importantly, the Lustre PFS - are all interconnected by an ultra-high speed network connection called [http://en.wikipedia.org/wiki/Infiniband InfiniBand]. A total of 7 InifiniBand switches are configured in a [http://en.wikipedia.org/wiki/Fat_tree fat tree] configuration.
| | |
| | == See also == |
| | |
| | * [[Compute Hardware Overview]] |
| | * [[Storage Systems Overview]] |
| | * [[Scheduler Overview (Slurm)]] |
| | * [[Log in to Anunna]] |
Anunna has a classic, heterogeneous HPC architecture: a set of login nodes that users connect to, a larger pool of compute nodes where the actual work runs, a fast parallel filesystem shared across the whole cluster, and a high-speed network tying it all together. The design is built to grow — nodes and storage can be added over time.
Login nodes
You reach the cluster by connecting to login.anunna.wur.nl over SSH (see Log in to Anunna). The login nodes are the entry point: you use them to edit files, manage your data, and submit jobs. They are shared by everyone, so they are not for heavy computation — anything beyond light, interactive work belongs in a job (see Batch Jobs and Interactive Jobs).
Compute nodes
The compute nodes do the actual work. They are heterogeneous: ordinary CPU nodes for most jobs, plus nodes equipped with NVIDIA or AMD GPUs for accelerated workloads. Jobs reach them through the SLURM scheduler, which places each job on a node that matches its requested resources.
- CPU work runs in the
main partition.
- GPU work runs in the
gpu (NVIDIA) and gpu_amd (AMD) partitions.
For the hardware details of each node type, see Compute Hardware Overview; for how to select a partition or constrain a job to particular hardware, see Partitions / Queues and Choosing a node (constraints).
Storage
A Lustre parallel filesystem (mounted at /lustre) is shared across all nodes and is the cluster's main working storage; your home directory lives on a separate, smaller NFS filesystem. Longer-term and archival storage is available too. For the full picture — quotas, backup behaviour, and which tier to use for what — see Storage Systems Overview.
Network
The nodes and the Lustre filesystem are connected by a high-speed, low-latency interconnect, which is what allows the parallel filesystem and multi-node jobs to perform well.
Operating system
The cluster runs Ubuntu 24.04 across its nodes. Software is provided through environment modules rather than being installed system-wide, so several versions of the same program can coexist.
See also
