Server and Cluster Requirements

Before installing and configuring AnzoGraph, you first need to determine the size and scale of the environment in which you want to deploy AnzoGraph, whether you are installing on "bare metal" servers or virtual machines, or using on-premises infrastructure or cloud resources.

The following sections detail the minimum requirements and recommendations to follow for setting up AnzoGraph host servers and cluster environments.

Server Hardware Requirements
Server Software Requirements
Firewall Requirements
Virtual Environments and Cluster Configuration

Server Hardware Requirements

Cambridge Semantics lists above average production system hardware requirements as a guideline. Large production data sets running interactive queries may require significantly more powerful hardware and RAM configurations. Provision production server hardware accordingly to avoid performance issues.

Component	Minimum	Recommended	Guidelines
Available RAM	16 GB (for small-scale testing only)	200 GB or more	AnzoGraph needs enough RAM to store data, intermediate query results, and run the server processes. Cambridge Semantics recommends that you allocate 3 to 4 times as much RAM as the planned data size. Do not overcommit RAM on a VM or on the hypervisor/container host. Avoid memory paging to disk (swapping) to achieve the highest possible level of performance. For more information about determining the server and cluster size that is ideal for hosting AnzoGraph, see Sizing Guidelines for In-Memory Storage.
Disk Space & Type	10 GB HDD	200+ GB SSD	AnzoGraph requires 10 GB for internal requirements. The amount of additional disk space required for load file staging and graph data persistence depends on the size of the data to be loaded. For persistence, Cambridge Semantics recommends that you have twice as much disk space on the local AnzoGraph file system as RAM on the server.
CPU	2	32	Once you provision sufficient RAM and a high-performing I/O subsystem, performance depends on raw CPU capabilities. Always use multi-core CPUs. A greater number of cores can make a dramatic difference in the performance of loads and queries. Intel processors are preferred, but AnzoGraph is supported on newer Epyc AMD processors. AnzoGraph does not run on older AMD processors.
Networking	10gbE	20+gbE	Not applicable for single server installations. Since AnzoGraph is high performance computing, massively parallel processing (MPP) graph OLAP engine, inter-cluster communications bandwidth dramatically affects performance. AnzoGraph clusters require optimal network bandwidth. All servers in a cluster must be in the same network. Make sure that all instances are in the same VLAN, security group, or placement group. In a switched network, make sure that all NICs link to the same Top Of Rack or Full-Crossbar Modular switch. If possible, enable SR-IOV and other HW acceleration methods and dedicated layer 2 networking that guarantees bandwidth.

AnzoGraph requires that all elements of the infrastructure provide the same quality of service (QoS). Do not run AnzoGraph on the same server as any other software except when in single-server mode and with an expectation of lowered performance. Providing the same QoS is especially important when using AnzoGraph in a clustered configuration. If any of the servers in the cluster perform additional processing, the cluster becomes unbalanced and may perform poorly. A single poor performing server degrades the other servers to the same performance level. All nodes require the same hardware specification and configuration. Also use static IP addresses or make sure that DHCP leases are persistent.

To ensure the maximum and most reliable QoS for CPU, memory, and network bandwidth, do not co-locate other virtual machines or containers (such as Docker containers) on the same hypervisor or container host. For hypervisor-managed VMs, configure the hypervisor to reserve the available memory for the AnzoGraph server. For clusters, make sure there is enough physical RAM to support all of the AnzoGraph servers, and reserve the memory via the hypervisor.

In addition, running memory compacting services such as Kernel Same-page Merging (KSM) impacts CPU QoS significantly and does not benefit AnzoGraph. Live migrations also impact the performance of VMs while they get migrated. While live migration can provide value for planned host maintenance, AnzoGraph performance may be impacted if live migrations occur frequently. For more information about Kernel Same-page Merging, see https://en.wikipedia.org/wiki/Kernel_same-page_merging.

Advanced configurations may benefit from CPU pinning on the hypervisor host and disabling CPU hyper-threading. For more information about CPU pinning, see https://en.wikipedia.org/wiki/Processor_affinity. For information about hyper-threading, see https://en.wikipedia.org/wiki/Hyper-threading.

Cambridge Semantics can provide benchmarks to establish relative cluster performance metrics and validate the environment.

Server Software Requirements

The table below lists the software requirements for AnzoGraph servers.

For container deployments, the required software and user resource limitation (ulimit) tuning is included in the AnzoGraph images. For RHEL/CentOS deployments, the instructions in Prerequisites and Pre-Installation Configuration provide details about configuring the required software and ulimits for standalone single server and cluster deployments.

Component	Requirement	Description
Operating System	RHEL/CentOS 7.6+	AnzoGraph is not supported on RHEL/CentOS 8 at this time.
GNU Compiler Collection	Installed	For compiling queries, AnzoGraph requires the latest version of the GCC tools for your operating system.

Optional Software

Program	Description
vim	Editor for creating or changing files.
sudo	Enables users to run programs with alternate security privileges.
net-tools	Networking utilities.
psutils	Python system and process utilities for retrieving information on running processes and system usage.
tuned	Linux system service to apply tunables.
wget	Utility for downloading files over a network.
Google SDK	For virtual servers on Google Cloud Engine (GCE). Command line tool to enable syncing of data from Google storage. You can download the latest version from Google: https://cloud.google.com/sdk/.

Firewall Requirements

AnzoGraph servers communicate via TCP/IP sockets. Client applications connect to AnzoGraph using the standard SPARQL HTTP(S) protocol. Cambridge Semantics applications, such as the AnzoGraph front end, communicate with the database via the secure, encrypted, gRPC-based protocol.

All servers in a cluster must be in the same network. Make sure that all instances are in the same VLAN, security group, or placement group.

Open the TCP ports listed in the table below. This image shows a visual representation of the communication ports:

Port	Description	Access needed...
5700	gRPC protocol port for secure communication between AnzoGraph servers.	Between all AnzoGraph servers in the cluster. Available for AnzoGraph on single node installations. Between the AnzoGraph leader server and any applications that connect to AnzoGraph using gRPC protocol, such as Apache Zeppelin and the AnzoGraph console.
8256	SPARQL HTTPS port for SSL communication between applications and AnzoGraph.	Between applications and the AnzoGraph leader server. Between all AnzoGraph servers in a cluster. Available for AnzoGraph on single node installations.
7070	SPARQL HTTP port for communication between applications and AnzoGraph.	Between applications and the AnzoGraph leader server. Between all AnzoGraph servers in a cluster. Available for AnzoGraph on single node installations.
9100	The internal fabric communications port.	Between all AnzoGraph servers in a cluster. Available for AnzoGraph on single node installations.
5600	The SSL system management port.	Between all AnzoGraph servers in a cluster. Available for AnzoGraph on single node installations.

Virtual Environments and Cluster Configuration

When your data loading and performance requirements warrant using a server cluster, the minimal size cluster you create should have no fewer than four nodes. When using a single node, data gets redistributed in memory without using a network. If you add one or two more nodes to create a two- or three-node cluster, data then gets distributed over the network to utilize the additional CPUs in the cluster. However, the CPU gain you would get from the additional one or two nodes added to the cluster does not outweigh the performance degradation that the network introduces.