Compute network requirements (NVMe/TCP)

CAUTION:

ANA setting must be enabled. If ANA is disabled, the expected I/O performance may not be achieved. In addition, the load on the internode network may increase beyond the expected level, causing delays in non-I/O processing within the VSP One SDS Block or resulting in processing failures.

The following table lists the compute network requirements to set up an NVMe/TCP connection configuration in VSP One SDS Block.

Item		Requirement	Remarks
Basic configuration		Configure as an L2 or L3 network. Configure as a separate network that does not have a connectivity with the internode network. To use the spare node function, the compute networks of the storage nodes and the spare node that belong to the same fault domain need to belong to the same subnet.	Use of a combination of NVMe/TCP and iSCSI or FC in the same storage cluster is not allowed. Configure a physical network that is different from the control network for performance stability and security improvement. When processing is switched to the spare node, the compute network information of the faulty storage node is passed to the spare node; thus, the compute networks of the storage nodes and the spare node that belong to the same fault domain need to belong to the same subnet. When multiple compute ports are set, and each of them makes up a separate subnet, the compute ports of the spare node to which the compute network information is passed also need to make up the same subnets. For the specific example, see Settings of spare node.
Connectivity	OS	Red Hat Enterprise Linux (64-bit) VMware ESXi	For details about the supported versions, contact customer support.
	NVMe/TCP access method	An NIC and NVMe over TCP software must be used.	When a combination of NIC and NVMe over TCP software is used, the NVMe/TCP communication performance viewed from a compute node depends largely on the CPU performance and memory capacity of the compute node. For details about the TCP/UDP port numbers to be used, see `TCP/UDP port numbers required for communication`. The maximum number of connections per session is 1.
	ANA setting	Asymmetric Namespace Access (ANA) of the compute node must be enabled.	The system might need to be restarted when the ANA setting is enabled.
Bandwidth		At least 1-Gbps bandwidth is required.	Perform operation with at least 10-Gbps bandwidth and jumbo frames (MTU = 9000 bytes) for secure performance.
Availability		When the network is redundant, the convergence time from a failure should be two minutes or less.	Make the network redundant so that it has no single point of failure (SPOF) that could disable communication in multiple storage nodes. An SPOF could cause a single network appliance failure or other link failure to stop I/O operations for the VSP One SDS Block cluster.
Security		Build an IP SAN completely independent of an external environment to prevent unauthorized access.	The In-band Authentication function is not supported.
Monitoring		Equipped with a network monitoring system independent of VSP One SDS Block to detect network failures and performance degradation (switch failures, link failures, increased delays, increased packet loss, connections at unexpected link speed operating values and duplex operating values, and others), and configure it so that the cause can be identified.	If the internode network is not provided with a monitoring system, it might be impossible to troubleshoot network failures or performance degradation.
Expandability		Determine a network topology and allocate IP addresses in consideration of the possibility that storage nodes might be added in the future.
Network topology		The network topology must be designed taking into account redundant configuration and future expansion. All compute ports are reachable from the compute node.	A configuration of two targets per storage node is recommended.
Maximum transfer length per I/O request		The transfer length per I/O request that reaches the volume from the compute node must be 256 KiB or less.	Typically, the compute node OS, driver, HBA, or hypervisor splits the transfer data so that the transfer data does not exceed the maximum transfer length per I/O request.