Factors to Consider for Capacity Planning
Choosing the right size of your cloud environment before migrating your workload from CDH to CDP Public Cloud is critical for preserving performance characteristics. There are several factors from your query workload to consider when choosing the CDP capacity for your environment.
- Query memory requirements
- CPU utilizations
- Disk bandwidth
- Working set size
- Concurrent query execution
Before getting into sizing specifics, it is important to understand the core hardware differences between PC and on-prem hosts:
|CDH Host Recommendation||AWS R5D.4xlarge Instance|
||128GB min. 256GB+ recommended||128|
||10Gbps min. 40Gbps recommended||Up to 10Gbps|
||12 x 2TB drives (1000MBps sequential)||2 x 300GB NVMe SSD (1100MBps sequential)|
||N/A||gp2 250 MB/s per volume|
An R5D.4xlarge instance closely matches the CDH recommended CPU, memory, and bandwidth specs and so it is recommended as the instance type for CDP. However, AWS ephemeral storage cannot be used as primary database storage since it is transient and lacks sufficient capacity. This core difference requires a different strategy for achieving good scan performance.
CDP Sizing and Scaling
Before migration, scaling and concurrency must be planned. In a public cloud environment, the ability to acquire better scaling and concurrency elastically in response to workload demand enables the system to operate at a lower cost than the maximum limits you plan for. If you configure your target environment to accommodate your peak workload as a constant default configuration, you might have cost overruns when system demand falls below that level.
In CDP, the T-Shirt size defines the number of executor instances for an individual cluster and hence determines memory limits and performance of individual queries. Conversely, the warehouse sizing parameter and auto-scaling determine how many clusters are allocated to support concurrent query execution.
|Sizes||Number of Executors|
The T-shirt size must be at least large enough to support the highest memory usage by a single query. In most cases, the size will not need to be larger but may provide better data caching if there is commonality in working sets between queries. Increasing the T-Shirt size can directly increase single-user capacity and also increase multi-user capacity. This is because the additional memory and resources from the larger cluster allows larger datasets to be processed and can also support concurrent query execution by sharing resources. Choosing a size that is too small can lead to poor data caching, spilling of intermediate results, or memory paging. Choosing a size this is too large can incur excessive PC run cost due to idle executors.
One caveat to consider when choosing a T-shirt size based on existing hardware is what other processes are running on the same host in your on-prem environment. In particular HDFS or other locally hosted filesystems may be consuming significant resources. You may be able to choose a smaller size for CDP since these processes will be isolated in their own pod in the CDP environment. It may be helpful to look at CM per-process metrics to isolate impalad and Impala frontend java processes that CDP will put on executor instances and to aggregate these metrics across your cluster. The java processes can accumulate significant memory usage due to metadata caching.
The size of your target environment corresponds to the peak concurrency the system can handle. Concurrency is the number of queries that can be run at the same time.
Each executor group can run 12 queries concurrently and occasional peaks can be handled transparently using the auto scaling feature. An autoscaling leading to add one more executor group doubles the query conncurrency to 24. Scaling the warehouse by adding more clusters allows for additional concurrent queries to run but will not improve single-user capacity or performance. This is because executors from the additional clusters are private to that cluster. Concurrently executed queries will be routed to the different clusters and run independently. The number of clusters can be changed to match concurrent usage by changing the autoscaling parameters.
Caching Hot Dataset
Currently CDH supports caching mechanisms on the compute nodes to cache the working set that is read from remote filesystems, such as, remote HDFS data node, S3, ABFS, and ADLS. This offsets the input/output performance difference.
In CDP Public Cloud, frequently accessed data is cached in a storage layer on SSD so that it can be quickly retrieved for subsequent queries, which boosts the performance. Each executor can have upto 200 GB of cache. So a medium-size can keep 200 * 20 = 4 TB of data in cache. For columnar formats (for example, ORC) data in cache is decompressed but not decoded. If the expected size of the hot dataset is 6 TB, which requires approximately 30 executors, you can choose to overprovision (choose a large) to ensure full cache coverage; or underprovision (choose a medium) to reduce cost at the expense of a lower rate of finding in the cache (cache hit rate).