Hardware requirements depending on workload patterns
The Hyperwave workloads present many resource-management challenges and potential conflicts. There are different types of conflicts, for example:
-
between long-running, resource-intensive jobs and shorter interactive queries;
-
between Hyperwave and non-Hyperwave workloads running on the same cluster and competing for the same resources.
In general, there are three main types of workload:
-
compute intensive;
-
storage intensive;
-
balanced.
Also, there could be situations, when you may not know your eventual workload patterns at first. Initial actions with an ADH cluster are usually very different from the actual jobs you will run in your production environment. Therefore, it is recommended following the advice for the balanced workload in a pilot ADH cluster. After that, you can plan or rescale the cluster according to the real workload.
Patterns
Same wheels
ADH nodes are like wheels that move everything. If the wheels are identical, then the movement is carried out smoothly and without jerks. But if they are different, then there may be different problems in the uniformity of movement. It is recommended to use the same configuration on all cluster machines with minor differences between DataNodes and between NameNodes.
The following characteristics must be the same for all cluster machines:
-
CPU;
-
RAM;
-
network.
See more details in General recommendations.
Compute intensive
This workload type is CPU-related and characterized by the need of a large number of CPU cores and large amount of memory to store in-process data. This usage pattern is typical for natural language processing or HPCC workloads. For the compute intensive patterns, it is necessary to use at least 10 CPU cores per machine.
Storage intensive
For this type of workload, it is recommended to invest in more disks per machine. The number of machines depends on the volume of data to store and analyze, which determines the required number of disks per machine. The latter is usually uniformal throughout the cluster.
For a low density server, the main aim is to keep the cost low to be able to afford a large number of machines. Eight CPU cores match this requirement and will give reasonable processing power. Each map or reduce task will utilize a single CPU core, but since some time will be spent waiting for I/O operations, it makes sense to oversubscribe for CPU cores. With 8 cores available, it is possible to configure about 12 map and reduce slots per node.
The optimal size of one hard disk is 2-3 TB. On average, you can install 12 disks in each machine.
Balanced workload
The logic behind choosing and balancing hardware components for a high density cluster is the same as for a lower density one. As an example of such a configuration, you can choose a machine with 12 2-3 TB hard drives or 24 1 TB hard drives. Having lower capacity disks per server is preferable, because it will provide better input and output throughput and better fault tolerance. To increase the computational power of an individual machine, use 8 CPU cores and 128-256 GB of RAM.
General recommendations
|
IMPORTANT
The following system requirements are minimal. The target sizing should be calculated based on the customer requirements. |
In this table, there are the most common hardware recommendations based on the workload pattern for an ADH cluster.
| Server | Load pattern | Storage | CPU | RAM | Network |
|---|---|---|---|---|---|
DataNodes |
Balanced workload |
12 Disks 2-3 TB JBOD |
8 cores |
128-256 GB |
1 Gbps onboard, 2x10 Gbps interconnect |
Compute-intensive workload |
12 Disks 1-2 TB JBOD |
10 cores |
128-256 GB |
10 Gbps onboard, 2x10 Gbps interconnect |
|
Storage-heavy workload |
12 Disks 4+ TB JBOD |
8 cores |
128-256 GB |
10 Gbps onboard, 2x10 Gbps interconnect |
|
NameNode |
Balanced workload |
4+ Disks 2-3 TB RAID 10 |
8 cores |
128-256 GB |
10 Gbps onboard, 2x10 Gbps interconnect |
Resource Manager/YARN |
Balanced workload |
4+ Disks 2-3 TB RAID 10 |
8 cores |
128-256 GB |
10 Gbps onboard, 2x10 Gbps interconnect |