Virtual Warehouse sizing requirements for public cloud environments

This topic describes how to estimate size requirements for Cloudera Data Warehouse (CDW) Public Cloud Virtual Warehouses.

Calculating public cloud requirements for on-premises data warehouse deployments

Selecting the correct size of public cloud environment before you migrate your workloads from CDH and HDP to Cloudera Data Warehouse (CDW) Public Cloud is critical for preserving performance characteristics. Consider the following workload characteristics when you plan for capacity on your public cloud environment:

  • Query memory requirements
  • CPU utilizations
  • Disk bandwidth
  • Working set size
  • Concurrent query execution requirements

As part of this calculation, it is important to understand the core hardware difference between public cloud and on-premises hosts as explained in the following table:

Table 1. Hardware differences between public cloud and on-premises hosts
Hardware component On-premises AWS R5D.4xlarge instance Azure E16 v3 instance
CPU cores 20 - 80 16 16
Memory

128 GB minimum

256 GB+ recommended

128 GB 128 GB
Network

10 GB/s minimum

40 GB/s recommended

Up to 10 GB/s Up to 8,000 MB/s
Instance storage

12 x 2 TB drives

(1,000 MB/s sequential)

2 x 300 GB NVMe SSD

(1,100 MB/s sequential)

400 GiB SSD
Persistent storage performance

At least 500 MB/s per disk

With 20 disks, 10 GB/s per node

1,156 MB/s per EC2 instance

Maximum IOPS: 24,000

Maximum Read: 375 MB/s

Maximum Write: 187 MB/s

An AWS R5D.4xlarge instance closely matches the CPU, memory, and bandwidth specifications that are recommended for CDH clusters. This is why the AWS R5D.4xlarge instance specifications were chosen as the "instance type" for CDP. That said, AWS EBS storage cannot be used as primary database storage because it is transient and lacks sufficient capacity. This core difference make it necessary to use a different strategy to achieve good scan performance.

Public cloud sizing and scaling

Before you migrate data to CDW Public Cloud, plan for scaling and concurrency. In the cloud, scaling and concurrency can elastically respond to workload demands, which enables the system to operate at a lower cost than you might expect. If you configure your environment to accommodate peak workloads as a constant default configuration, you may waste resources and money when system demand falls below that level.

In CDW, the size of the Virtual Warehouse determines the number of executor instances for an individual cluster so it determines memory limits and performance capabilities of individual queries:

Table 2. Virtual Warehouse sizes
Size Number of executors
X-Small 2
Small 10
Medium 20
Large 40

Warehouse size in combination with auto-scaling settings determine how many clusters are allocated to support concurrent query execution.

The Virtual Warehouse size must be at least large enough to support the memory used by the most data-intensive query. Usually, the Virtual Warehouse size does not need to be larger that the most data-intensive query. Better caching is provided if there is commonality between data sets accessed by queries. Increasing the Virtual Warehouse size can increase single-user and multi-user capacity because additional memory and resources allows larger datasets to be processed. Concurrent query execution is also supported by sharing resources. If too small a size is configured for the Virtual Warehouse, poor data caching and memory paging can result. If too large a size is configured, excessive public cloud costs are incurred due to idle executors.

The primary difference between CDW Public Cloud and CDH on-premises deployments when choosing a warehouse size based on existing hardware:
  • With CDW Public Cloud all resources on an executor are dedicated to query processing.
  • With CDH on-premises deployments resources support other operations in addition to query processing. For example, these on-premises resources are shared with other services, such as HDFS or other locally hosted file systems. In particular Spark, HBase, or MapReduce. These other services might consume significant resources.
Consequently, you might be able to choose a much smaller Virtual Warehouse size in CDW Public Cloud because resources are isolated in their own pod in the CDW Public Cloud environment.

In the case of Impala, it is useful to look at the Cloudera Manager per-process metrics to isolate the impalad backend and the Impala front-end Java processes that hold the Catalog cache. In CDW Public Cloud, the Impala coordinator and executor roles are separated leaving the unused Catalog JVM memory free to support query execution. You should look at the memory utilization metrics for executor-only impalad nodes (those not also running the coordinator role) to estimate how much memory your current cluster of Impala executors requires.

Concurrency

Concurrency is the number of queries that can be run at the same time. Determine the size you need by considering the amount of resources your system needs to support peak concurrency.

For Impala Virtual Warehouses, each executor group can run 12 queries concurrently. Occasional peaks are handled transparently by the auto-scaling feature. When auto-scaling is triggered an additional executor group is added thereby doubling query concurrency capacity to 24. Scaling the Virtual Warehouse by adding more clusters enables additional concurrent queries to run, but does not improve single-user capacity or performance. Concurrently executed queries are routed to the different clusters and execute independently. The number of clusters can be changed to match concurrent usage by changing the auto-scaling parameters. For more details about auto-scaling settings, see the links at the bottom of this page.

For Hive on LLAP Virtual Warehouses, each size setting indicates the number of concurrent queries that can be run. For example, an X-Small Hive on LLAP Virtual Warehouse can run 2 concurrent queries. A Small Virtual Warehouse can run 10 concurrent queries. To run 20 concurrent queries in a Hive on LLAP Virtual Warehouse choose Medium size.

Caching "Hot Datasets"

Frequently accessed data is sometimes referred to as a "hot dataset." Currently, CDH supports caching mechanisms on the compute nodes to cache the working set that is read from remote file systems, such as remote HDFS data nodes, S3, ABFS, or ADLS. This offsets the input/output performance difference.

In CDW Public Cloud, frequently accessed data is cached in a storage layer on SSD so that it can be quickly retrieved for subsequent queries. This boosts performance. Each executor can have up to 200 GB of cache. For example, a Medium-sized Virtual Warehouse can keep 200 * 20 = 4 TB of data in its cache. For columnar formats, such as ORC, data in the cache is decompressed, but not decoded. If the expected size of the hot dataset is 6 TB, which requires about 30 executors, you can over-provision by choosing a Large-sized warehouse to ensure full cache coverage. A case can also be made to under-provision by choosing a Medium-sized warehouse to reduce costs at the expense of having a lower cache hit rate. To offset this, keep in mind that columnar formats allow optimizations such as column projection and predicate push-down, which can significantly reduce cache requirements. In these cases, under-provisioning the cache might have no negative performance effect.

Scanned dataset size

Scanning large datasets on Amazon S3 or Azure ADLS can be slow. A single R5D.4xlarge EC2 instance can only scan data at 1,156 MB/s maximum throughput according to standard S3 benchmarks. If a query must read 100 GB from S3, S3 scanning takes a minimum of 88 seconds on just one node. Depending on the number of files in the S3 directory, it might take more than 3 minutes. In this case, if a query needs to scan 100 GB of data, if you use 10 nodes, you can get the scan time down to approximately 20 seconds.

However, keep in mind that with columnar storage, minimum/maximum statistics in files, and other conditions, often the amount of data read is significantly less than the total size of the files for certain queries.