| 1 |
Performance analysis with Firebird. |
| 2 |
Document version 0.2 |
| 3 |
Created by Nickolay Samofatov |
| 4 |
|
| 5 |
Performance monitoring framework is consisted of 2 parts: |
| 6 |
- Trace API |
| 7 |
- Reference implementation of analysis plugin |
| 8 |
|
| 9 |
|
| 10 |
Trace API. |
| 11 |
|
| 12 |
Trace API allows to plug a monitoring module into the engine and notify it of |
| 13 |
processes happening there. List of analysis plugins is specified in firebird.conf |
| 14 |
parameter TracePlugins. Each trace plugin receives notifications on each database |
| 15 |
opened by the engine as a call of procedure having following signature: |
| 16 |
|
| 17 |
NTRACE_BOOLEAN trace_attach(const char* db_filename, const TracePlugin** plugin); |
| 18 |
|
| 19 |
Conceptually TracePlugin is the plugin trace interface. It is implemented as a |
| 20 |
structure of function pointers and void* object pointer to be passed as first |
| 21 |
argument for each function. |
| 22 |
|
| 23 |
Given this interface plugin may install hooks for particular events and receive |
| 24 |
notifications for them. If plugin has interest in some kind of events for given |
| 25 |
database it sets plugin pointer to a structure containing non-null hook function |
| 26 |
pointers and returns NTRACE_TRUE value. Otherwise it should return NTRACE_FALSE. |
| 27 |
|
| 28 |
One category of hooks receives the same parameters as corresponding Y-Valve |
| 29 |
(public API) methods plus internal objects identifiers, generated PLANs for |
| 30 |
statements and performance counters. One more hook receives performance |
| 31 |
statistics for each procedure execution. |
| 32 |
|
| 33 |
Each hook function returns boolean value indicating success or failure. When hook |
| 34 |
function fails engine collects error message from plugin in the same thread |
| 35 |
context as hook was called, writes message to firebird.log and stops calling |
| 36 |
hooks for this particular database and trace plugin. |
| 37 |
|
| 38 |
See jrd/ntrace.h module for definitions of structures and events supported. |
| 39 |
|
| 40 |
|
| 41 |
Default analysis module. |
| 42 |
|
| 43 |
Reference implementation of analysis plugin is centered around a concept of |
| 44 |
performance snapshot. Snapshot detail may vary, but in any case it contains |
| 45 |
information on all activities performed by server for database during a given |
| 46 |
period of time, possibly in aggregated form. All snapshots may be compared to |
| 47 |
allow quantitative measurements of performance improvements due to various |
| 48 |
optimization activities or changes in user applications and infrastucture. |
| 49 |
|
| 50 |
Snapshot data is collected by trace module in raw form in a binary log file and |
| 51 |
then may be loaded into a stats database using ntrace utility. |
| 52 |
|
| 53 |
Stored snapshots may contain data on usage of indices, statements executed, their |
| 54 |
plans and performance counters, IO statistics and other performance data. |
| 55 |
|
| 56 |
Level of detail for information collected is set up in analysis trace module |
| 57 |
configuration file on per-database basis. |
| 58 |
|
| 59 |
Default analysis module does not have GUI and only provides a database with data |
| 60 |
on server activities and some stored procedures to use and maintain it. |
| 61 |
|
| 62 |
It is expected that third parties develop GUI tools, custom trace modules and |
| 63 |
statistics tools to statisfy rich client needs. |