From 11eedc580395ff297a7371b49c07e66fc05a6038 Mon Sep 17 00:00:00 2001 From: schmidt-scaled <82834682+schmidt-scaled@users.noreply.github.com> Date: Fri, 19 Jun 2026 16:14:41 +0300 Subject: [PATCH 1/2] Update numa-considerations.md --- .../deployment-preparation/numa-considerations.md | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/docs/deployments/deployment-preparation/numa-considerations.md b/docs/deployments/deployment-preparation/numa-considerations.md index 08150219..7778c883 100644 --- a/docs/deployments/deployment-preparation/numa-considerations.md +++ b/docs/deployments/deployment-preparation/numa-considerations.md @@ -9,12 +9,12 @@ Modern multi-socket servers use a memory architecture called In a NUMA system, each CPU socket has its own local memory and I/O paths. Accessing local resources is faster than reaching across sockets to remote memory or devices. Simplyblock is fully NUMA-aware. -On a host with more than one socket, by default one or two storage nodes are deployed per socket. +On a host with more than one socket, by default one storage node is deployed per socket. -Two storage nodes per socket are deployed if: +In rare cases, it can make sense to deploy more than one node per socket (disaggregated deployment only): -- more than 32 vCPUs (cores) per NUMA socket are dedicated to simplyblock per socket -- more than 10 NVMe devices are connected to the NUMA socket +- more than 10 NVMe attached to a single socket +- more than 48 cores (96 vcpu) on a single socket Users can change this behavior. Either by setting the appropriate Helm Chart parameters (in case of Kubernetes-based storage node deployment) or by manually modifying the initially created configuration file on the storage node @@ -28,6 +28,10 @@ If a socket has no NVMe devices connected, it will not qualify to run a simplybl It is also important that the NIC(s) used by simplyblock for storage traffic are connected to the same NUMA socket. However, simplyblock does not auto-assign a NIC and users have manually to take care of that. +!!! Important + In a hyper-converged deployment, it can be preferable to attach all NVMe to one socket and only deploy a node on + this socket. This saves overall CPU and RAM required for Simplyblock. + ## Checking NUMA Configuration Before configuring simplyblock, the system configuration should be checked for multiple NUMA nodes. This can be done From 26dff8c7c4faa689f97413d22caa9a6e6bfbd878 Mon Sep 17 00:00:00 2001 From: noctarius aka Christoph Engelbert Date: Fri, 19 Jun 2026 15:18:51 +0200 Subject: [PATCH 2/2] Update NUMA terminology in documentation --- .../deployment-preparation/numa-considerations.md | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/docs/deployments/deployment-preparation/numa-considerations.md b/docs/deployments/deployment-preparation/numa-considerations.md index 7778c883..a1e48f6c 100644 --- a/docs/deployments/deployment-preparation/numa-considerations.md +++ b/docs/deployments/deployment-preparation/numa-considerations.md @@ -9,12 +9,12 @@ Modern multi-socket servers use a memory architecture called In a NUMA system, each CPU socket has its own local memory and I/O paths. Accessing local resources is faster than reaching across sockets to remote memory or devices. Simplyblock is fully NUMA-aware. -On a host with more than one socket, by default one storage node is deployed per socket. +On a host with more than one socket, by default one storage node is deployed per NUMA node. -In rare cases, it can make sense to deploy more than one node per socket (disaggregated deployment only): +In rare cases, it can make sense to deploy more than one storage node per NUMA node (disaggregated deployment only): -- more than 10 NVMe attached to a single socket -- more than 48 cores (96 vcpu) on a single socket +- A single socket has attached more than 10 NVMe. +- A single socket provides more than 48 cores (96 vCPUs). Users can change this behavior. Either by setting the appropriate Helm Chart parameters (in case of Kubernetes-based storage node deployment) or by manually modifying the initially created configuration file on the storage node @@ -29,8 +29,8 @@ It is also important that the NIC(s) used by simplyblock for storage traffic are However, simplyblock does not auto-assign a NIC and users have manually to take care of that. !!! Important - In a hyper-converged deployment, it can be preferable to attach all NVMe to one socket and only deploy a node on - this socket. This saves overall CPU and RAM required for Simplyblock. + In a hyper-converged deployment, it can be preferable to attach all NVMe devices to one socket and only deploy a + node on this socket. This saves overall CPU and RAM required for Simplyblock. ## Checking NUMA Configuration