Planning a Weka System Installation
Last updated
Last updated
The planning of a Weka system is essential prior to the actual installation process. It involves the planning of the following:
Total SSD net capacity and performance requirements
SSD resources
Memory resources
CPU resources
Network
A Weka system cluster runs on a group of hosts with local SSDs. To plan these hosts, the following information must be clarified and defined:
Capacity: Plan your net SSD capacity. Note that data management to object stores can be added after the installation. In the context of the planning stage, only the SSD capacity is required.
Redundancy scheme: Define the optimal redundancy scheme required for the Weka system, as explained in .
Failure domains: Determine whether failure domains are going to be used (this is optional) and if yes determine the number of failure domains and the potential number of hosts in each failure domain, as described in , and plan accordingly.
Hot spare: Define the required hot spare count, as described in .
Once all this data is clarified, you can plan the SSD net storage capacity accordingly, as defined in the . You should also have the following information which will be used during the installation process:
Cluster size (number of hosts).
SSD capacity for each host, e.g., 12 hosts with a capacity of 6 TB each.
Planned protection scheme, e.g., 6+2.
Planned failure domains (optional).
Planned hot spare.
SSD resource planning involves how the defined capacity is going to be implemented for the SSDs. For each host, the following has to be determined:
The number of SSDs and capacity for each SSD (where the multiplication of the two should satisfy the required capacity per host).
The technology to be used (NVME, SAS, or SATA) and the specific SSD models, which have implications on SSD endurance and performance.
The total per host memory requirements is the sum of the following requirements:
Purpose
Per host memory
Fixed
2.3 GB
Frontend cores
2.3 GB x # of Frontend cores
Compute cores
3.3 GB x # of Compute cores
Drive cores
2.3 GB x # of Drive cores
SSD capacity management
HostSSDSize/10,000 (HostSSDSize = Total SSD raw capacity / # of hosts)
Operating System
The maximum between 8 GB and 2% from the total RAM
Additional protocols (NFS/SMB/S3)
8 GB
RDMA
2 GB
Metadata (pointers)
A system with 16 hosts with the following details:
Number of Frontend cores: 1
Number of Compute cores: 13
Number of Drive cores: 6
Total raw capacity: 983 TB
Total net capacity: 725 TB
NFS/SMB services
RDMA
Average file size: 1 MB (potentially up to 755 million files for all hosts; ~47 million files per host)
Calculations:
Frontend cores: 1 x 2.3 = 2.3 GB
Compute cores: 13 x 3.3 = 33.9 GB
Drive cores: 6 x 2.3 = 13.8 GB
SSD capacity management: 983 TB / 16 / 10K = ~6.3 GB
Metadata: 20 Bytes x 47 million files x 2 units = ~1.9 GB
Total memory requirement per host = 2.3 + 2.3 + 33.9 + 13.8 + 6.3 + 8 + 2 + 1.9 = ~71 GB
For the same system as in example 1, but with smaller files, the required memory for metadata would be larger.
For an average file size of 64 KB, the number of files is potentially up to ~12 billion files for all hosts; ~980 million files per host.
Required memory for metadata: 20 Bytes x 980 million files x 1 unit = ~19.6 GB
Total memory requirement per host = 2.3 + 2.3 + 33.9 + 13.8 + 6.3 + 8 + 2 + 19.6 = ~88 GB
The Weka software on a client host requires 4 GB of additional memory.
The Weka system implements a Non-Uniform Memory Access (NUMA) aware CPU allocation strategy to maximize the overall performance of the system. The allocation of cores utilizes all NUMAs equally to balance memory usage from all NUMA nodes.
The following should be noted with regards to the CPU allocation strategy:
The code allocates CPU resources by assigning individual cores to tasks in a cgroup
Cores in a Weka cgroup won't be available to run any other user processes
On systems with Intel hyperthreading enabled, the corresponding sibling cores will be placed into a cgroup along with the physical ones.
The number of physical cores dedicated to the Weka software should be planned according to the following guidelines and limitations:
At least one physical core should be dedicated to the operating system; the rest can be allocated to the Weka software.
In general, it is recommended to allocate as many cores as possible to the Weka system.
No more than 19 physical cores can be assigned to Weka system processes.
Enough cores should be allocated to support performance targets.
In general, use 1 drive core per SSD for up to 6 SSDs and 1 drive core per 2 SSDs for more, with a ratio of 2 compute cores per SSD core.
For finer tuning, please contact the Weka Support Team.
Enough memory should be allocated to match core allocation, as discussed above.
The running of other applications on the same host (converged Weka system deployment) is supported. However, this is not covered in this documentation. For further information, contact the Weka Support Team.
On a client host, by default, the Weka software consumes a single physical core. If the client host is configured with hyper-threading, the Weka software will consume two logical cores.
If the client networking is defined as based on UDP, there is no allocation of core resources and the CPU resources are allocated to the Weka processes by the operating system as any other process.
WEKA backend servers support connections to both InfiniBand and Ethernet networks. When deploying backend servers, ensure that all servers in the WEKA system are connected using the same network technology for each type of network.
If both InfiniBand and Ethernet connections are configured, the WEKA system prioritizes InfiniBand links for data traffic. However, if there is a connectivity issue with the InfiniBand network, the system automatically switches to using Ethernet links as a fallback. Clients can connect to the WEKA system over either InfiniBand or Ethernet.
Clients can be configured with networking as described above to achieve the highest performance and lowest latency; however, this setup requires compatible hardware and dedicated CPU core resources. If compatible hardware is not available or a dedicated CPU core cannot be allocated to the WEKA system, client networking can instead be configured to use the kernel’s UDP service. This configuration results in reduced performance and increased latency.
20 Bytes x # Metadata units per host See .
Note: The memory requirements are conservative and can be reduced in some situations, such as in systems with mostly large files or a system with files 4 KB in size. Contact the to receive an estimate for your specific configuration.