Worker Allocation on Engine Yard Cloud
Workers are the processes that allow your application to respond to incoming web requests. Regardless of what Application Server Stack you select, it will use one or more workers per instance to run your application.
Standard Allocations based on Instance vCPU and Memory Configuration
For Passenger, Unicorn, Thin or Mongrel, these workers are allocated using an algorithm that takes into consideration vCPU and Memory available to the instance to determine the number of workers to allocate, as shown in the table below.
We are also happy to announce we have recently modified the algorithm that generates the worker counts to better make use of the available resources, which has increased the number of workers for most sizes.
|Type||Memory (MB)||Swap (MB)||vCPUs||Pool Size|
|Original||New Solo||New Cluster|
|High CPU Medium||1740||1343||5||6||6||6|
|High CPU Extra Large||7168||30725||20||24||40||40|
|High Memory Extra Large||17500||0||6.5||8||13||13|
|High Memory Double Extra Large||35000||0||13||8||26||26|
|High Memory Quad Extra Large||70000||30725||26||24||52||52|
|High I/O Quad Extra Large||62000||30725||35||70||70||70|
Note about vCPUs: vCPUs is a generic label for virtual CPU cores (and in some cased hyperthreads) available to your instance. Amazon, the provider of the instances used on Engine Yard Cloud, uses the term ECU, or Elastic Compute Units, to convey the relative compute power available on each instance type. This value is different than the virtual CPUs reported by the operating system, such as in top or /proc/cpuinfo, but better allows you to determine how much effective compute resources you need. According to Amazon, the definition of ECU is:
One EC2 Compute Unit provides the equivalent CPU capacity of a 1.0-1.2 GHz 2007 Opteron or 2007 Xeon processor.
Tuning for your Application's Needs
The values above are based on generic parameters of an idealized application. However, if your app falls outside of these parameters, you should consider tuning your environment, to either:
- reduce the worker counts to improve responsiveness because your app requires more resources
- increase available workers for a lightweight app with high requests per minute
Tuning is done by setting parameters on your environment. The sections below will explain the various parameters that can be tuned. Once you know what values you want tuned, you can open a support ticket to have them put in effect.
In the case of Medium and High CPU Medium instances, there in an override in place to ensure they didn't reduce their pool size from the values provided in the original algorithm. If you tune one or more parameters, these overrides are removed and the generated values based on those parameters will be used instead.
This is the most direct way of modifying your pool size. By specifying
pool_size, you can set it regardless of resource availability. This is
recommended only if you are fully aware of your resource needs and don't plan
on changing your instance size in the future.
Slightly less forceful, the
min_pool_size parameter allows you to ensure
your application always has the minimum number of workers available. Useful
if using the smaller instance types and you don't mind overloading the
resources -- be careful though, overloaded instances may result in
underperforming or even hung instances, potentially triggering a takeover.
If you find yourself using this and having poor responsiveness or takeovers, please consider upgrading to a larger instance size.
max_pool_size parameter allows you to ensure you don't have
more workers than you need, regardless of resource availability. You may want
to do this if your workers connect to external entities that cannot handle too
many simultaneous connections.
This is the best and most accurate way of tuning your app based on memory usage. The standardized values above are based on an app in which each worker uses 250MB of memory. If your app is significantly different than this, it is best to tune this value to get accurate resource usage calculations.
Some apps are compute heavy, others barely use the any processor capacity. The standard configuration provides 2 workers per ECU; tuning this to your app's needs can improve its responsiveness.
While it is not a good idea to go in to swap in most cases, it is a balancing act to provide the optimal number of workers. We have chosen a modest 25% usage of swap by the workers in order to provide a sufficient worker count without excessive paging. If you find that your app is spending a great deal of time in swap, you may want to change this value.
Note: It is usually due to CPU constraints rather than memory constrants that determines the workers available, but depending on how many workers you allow per CPU, this may not be the case. Under default settings, this value most significanly impacts the Small, Medium, and High CPU Extra Large instances.
Default: 25 (%)
If you are running background workers or other processes that eat into your memory resources, you may want to specify a chunk of memory to reserve. By default, this is set to a value that accomodates the OS and support processes.
Default: 1500 (MB)
reserved_memory but also takes into consideration the database
that shared the instance in a single instance (aka Solo) environment.
Default: 2000 (MB)
db_vcpu_max and db_workers_per_vcpu
These settings are only applicable to single instance environments. By
default, no consideration is taken to reserve ECU resources for the database,
but if you have a database-heavy application, it would be beneficial to
allocate some of the worker resources to the database instead. The number of
workers represented by the
db_workers_per_vcpu parameter is removed from
each ECU, up to the
db_vcpu_max = 0;
db_workers_per_vcpu = 0.5
What about other Application Server Stacks?
Concurrent application servers, such as Puma (all Ruby Runtimes) and Trinidad (JRuby only), use a simple one worker per ECU rule, and rely on threading to scale to the incoming requests. These work best for concurrent Ruby implementations such as Rubinius and JRuby, but threadsafe apps properly configured can benefit from concurrency under MRI by using Puma.