Alexandr created GUACAMOLE-2318:
-----------------------------------
Summary: Reduce write() overhead when encoding Guacamole protocol
instructions in libguac
Key: GUACAMOLE-2318
URL: https://issues.apache.org/jira/browse/GUACAMOLE-2318
Project: Guacamole
Issue Type: Improvement
Components: guacd
Reporter: Alexandr
Right now every field of a guac_protocol_send_* instruction (copy, transform,
cursor, etc.) goes to the socket as its own separate write - for something like
copy that's 30+ small locked writes per instruction, each taking the buffer
mutex and doing its own memcpy, plus a redundant snprintf per field.
Batched each "<length>.<value>" element into a single stack buffer and write it
in one call instead. Wire format is unchanged - this only cuts the number of
write() calls.
Benchmarked against a real build Linux/glibc:
{code:java}
copy: ~1730 → ~1220 ns/call (-30%)
transform: ~2020 → ~1610 ns/call (-20%)
name: ~227 → ~205 ns/call (-10%){code}
Pure CPU win, no change in memory usage - same stack buffers, no new
allocations.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
