Skip to content

Latest commit

 

History

History
43 lines (33 loc) · 2.29 KB

performance.md

File metadata and controls

43 lines (33 loc) · 2.29 KB
Raw

Performance

zzz's design philosophy results in a lot of knobs that the consumer of the library can turn and tune to their preference.

These performance tips are general and can apply to any protocol implementation. HTTP is used as the general example because it is currently the only completed protocol.

Performance Hunting

zzz now officially runs multithreaded by default. By default, it will utilize @min(cpu_count / 2 - 1, 1) threads. This can be tuned by changing the .threading option of the Tardy runtime.

var t = try Tardy.init(.{
    .allocator = allocator,
    .threading = .{ .multi = COUNT },
});

The most important part of switching to the multi threaded model is using a thread-safe allocator. The general purpose allocator with the thread safe flag is an ideal choice.

Another way to alter the performance is by changing which Async I/O model you use. .auto will select the best one overall but there are other options that may be worth benchmarking depending on your use case.

Other settings of note include:

  • connection_arena_bytes_retain which controls how much memory that has been allocated within a connection's arena will be retained for the next connection.
  • connections_size_max which controls the maximum number of connections that each thread can handle. Any connection after this number will get closed.

Minimizing Memory Usage

When using zzz in certain environments, your goal may be to reduce memory usage. zzz provides a variety of controls for handling how much memory is allocated at start up.

var server = Server.init(rt.allocator, .{
    .backlog_count = 32,
    .connection_count_max = 16,
    .connections_size_max = 16,
    .connection_arena_bytes_retain = 64,
    .socket_buffer_bytes = 512,
});

There is no overarching setting here but a selection of ones you can tune to minimize:

  • run in single threaded mode
  • connections_size_max can be reduced. This value is used internally as every connection on every thread owns a Provision.
  • connection_arena_bytes_retain can be reduced. If you are not doing any allocations within the handler, you can even set this to 0.
  • socket_buffer_bytes can be reduced. There is a lower limit to this value of around 128 bytes. (This will be changed in the future and internal buffers will be separated).