Analytics In Real-time (AIR) is a light-weight system profiling tool that provides a set of APIs for profiling performance, latency, histogram, and so on.
For now, AIR supports Linux system only. This is an AIR tutorial to profile an application. Here is 5 Steps to use AIR:
- Update AIR configuration
- Build AIR with CMake
- Insert AIR APIs to an application
- Build an application with AIR
- Visualize AIR data
For more details, see AIR User Guide documentation.
First of all, user needs to define a configuration to use AIR APIs. This configuration includes AIR system wide rules and node specific information. This configuration will be interpreted at compile-time. Therefore AIR could validate syntax and optimize profiling code at compile-time.
[DEFAULT]
"AirBuild: True, StreamingInterval: 3, NodeBuild: True, NodeRun: On, NodeSamplingRatio: 1000, NodeIndexSize: 10"
[/DEFAULT]
[GROUP]
"Group:SUBMIT, NodeIndexSize: 5"
"Group:COMMON"
[/GROUP]
[FILTER]
"Filter: AIR_IOtype, Item: (AIR_READ, AIR_WRITE)"
"Filter: AIR_Thread, Item: (AIR_SUBMIT, AIR_PROCESS, AIR_COMPLETE)"
"Filter: AIR_Range, Item: (AIR_0 ... AIR_3)"
[/FILTER]
[BUCKET]
"Bucket: BUCKET_1, Bounds: [0, 100), Scale: 10"
"Bucket: BUCKET_4, Bounds: [2^0, 2^10), Scale: 2^"
[/BUCKET]
[NODE]
"Node: PERF_BENCHMARK, Filter: AIR_IOtype, Type: Performance, Group: COMMON"
"Node: LAT_PROCESS, Filter: AIR_Range, Type: Latency, Group: COMMON"
"Node: LAT_IO_PATH, Filter: AIR_Range, Type: Latency, Group: COMMON, IndexSize: 3"
"Node: UTIL_SUBMIT_THR, Filter: AIR_Thread, Type: Utilization, Group: SUBMIT, Build: False"
"Node: CNT_TEST_EVENT, Filter: AIR_Thread, Type: Count, Group: SUBMIT, IndexSize: 10"
"Node: HIST_SAMPLE_1, Filter: AIR_Range, Type: Histogram, Group: COMMON, IndexSize: 10, Bucket: BUCKET_1"
"Node: HIST_SAMPLE_4, Filter: AIR_Range, Type: Histogram, Group: COMMON, IndexSize: 10, Bucket: BUCKET_4"
[/NODE]
For more details, see Test Config file.
When the configuration file is updated, user has to build AIR to adjust the changes.
git clone https://github.com/poseidonos/air.git
cd air
mkdir build
cd build
cmake ..
If you want to build with your own config file, you should replace the last command with
cmake .. -DCONFIG=/your_file_directory/filename
After cmake, you can see a Makefile in the build directory. Type 'make' to build AIR. And then you can install AIR in your system if you are a system administrator.
make
sudo make install
If you want to build AIR with other libraries, you can use ExternalProject_Add option.
ExternalProject_Add(air
#---Download step-----------------
GIT_REPOSITORY ${REPO_AIR}
GIT_TAG ${TAG_AIR}
#---Update/Patch step-------------
UPDATE_COMMAND mkdir -p ${DIR_AIR}/build
#---Configure step----------------
SOURCE_DIR ${DIR_AIR}
CONFIGURE_COMMAND cmake .. -DCONFIG=${CFG_AIR}
#---Build step--------------------
BINARY_DIR ${DIR_AIR}/build
BUILD_COMMAND make
#---Install step------------------
INSTALL_COMMAND make install
#---Test step---------------------
#---Output logging----------------
LOG_DOWNLOAD ON
LOG_UPDATE ON
LOG_PATCH ON
LOG_CONFIGURE ON
LOG_BUILD ON
LOG_INSTALL ON
LOG_OUTPUT_ON_FAILURE ON
LOG_MERGED_STDOUTERR ON
)
AIR has a number of APIs which user insert by oneself. If user wants to profile some point of code, user needs to insert a proper API. For more details, see AIR User Guide documentation.
AIR has a polling thread which deals with raw data processing, air_cli response, JSON format output streaming. To run this thread, user should insert some APIs below.
...
#include <air/Air.h>
...
int main(int argc, char *argv[])
{
...
air_initialize(profiling_core); // AIR running(pinning) core, default: 0
air_activate();
// application code
...
air_deactivate();
air_finalize();
return retVal;
}
...
When AIR collects data, it can classify thread ID where the data came from. User doesn't need to concern about thread ID. AIR automatically handles it.
...
#include <air/Air.h>
...
void
AIO::SubmitAsyncIO(pos_io& posIo)
{
...
case UbioDir::Write:
{
airlog("PERF_VOLUME", "AIR_WRITE", posIo.volume_id, posIo.length);
...
}
case UbioDir::Read:
{
airlog("PERF_VOLUME", "AIR_READ", posIo.volume_id, posIo.length);
...
}
...
}
The above airlog(...) codes are C++ style APIs for performance profiling. Through this APIs, AIR can get the IOPS, Bandwidth per volume & per thread.
...
#include <air/Air_c.h>
...
static int _bdev_pos_eventq_rw(struct spdk_io_channel *ch, struct spdk_bdev_io *bdev_io){
...
case SPDK_BDEV_IO_TYPE_READ:
AIRLOG(LAT_BDEV_READ, BDEV_IO_BEGIN, vol_id, (uint64_t)bdev_io);
...
case SPDK_BDEV_IO_TYPE_WRITE:
AIRLOG(LAT_BDEV_WRITE, BDEV_IO_BEGIN, vol_id, (uint64_t)bdev_io);
...
}
...
static void bdev_pos_io_complete(struct pos_io* io, int status)
{
...
if (READ == io->ioType) {
AIRLOG(LAT_BDEV_READ, BDEV_IO_END, io->volume_id, (uint64_t)io->context);
}
else if (WRITE == io->ioType) {
AIRLOG(LAT_BDEV_WRITE, BDEV_IO_END, io->volume_id, (uint64_t)io->context);
}
...
}
The above AIRLOG(...) codes are C style APIs for latency profiling. Through this APIs, AIR can get some statistics such as min/avg/max value of bdev layer's IO latency.
After AIR build, user can link libair.a library to their application
target_link_libraries(${target_application} air ...)
LDFLAGS += -L/usr/local/lib -lair
There is a way to visualize AIR data.
TUI is a primitive and direct way. User can see the result on the machine where the application run. TUI displays the latest time result. To run TUI, just execute air_tui on the terminal.
TUI mode: [online], status: [play], page: [6/6], interval: [3], timestamp: 2022-3-4:11:12:2, pid: 27019
key {↑/↓}: move, {→/←}: (un)fold, {b}: perv page, {n}: next page, {space}: play/pause
{1~9}: streaming interval, {i}: init, {o}: run, {x}: stop, {q(esc)}: quit
(*) Top
( ) +Group:COMMON
( )[O]++Node:LAT_IO_PATH("latency")
""(0), index:0x0, filter:"AIR_0~AIR_1" Period(avg:262ns, median:126ns, max:5.7us, sample:99 ), Cumulation(avg:507ns, median:188ns, max:95us , sample:499 )
""(0), index:0x0, filter:"AIR_1~AIR_2" Period(avg:224us, median:222us, max:233us, sample:54 ), Cumulation(avg:224us, median:222us, max:233us, sample:54 )
""(0), index:0x1, filter:"AIR_0~AIR_1" Period(avg:224ns, median:140ns, max:1.1us, sample:94 ), Cumulation(avg:259ns, median:151ns, max:6.5us, sample:479 )
( )[O]++Node:LAT_PROCESS("latency")
""(0), index:0x200, filter:"AIR_0~AIR_1" Period(avg:497ns, median:361ns, max:1.6us, sample:99 ), Cumulation(avg:685ns, median:426ns, max:9.7us, sample:497 )
""(0), index:0x201, filter:"AIR_0~AIR_1" Period(avg:587ns, median:436ns, max:4.1us, sample:100 ), Cumulation(avg:780ns, median:533ns, max:9.8us, sample:498 )
( )[O]++Node:PERF_BENCHMARK("performance")
SUM Period(iops:2.1M , bw:8.8GB ), Cumulation(iops:2.1M , bw:8.7GB )
"CompleteIO"(27027), index:0x200, filter:"AIR_READ" Period(iops:1.1M , bw:4.4GB , count:[{"4096": 3609331}]), Cumulation(iops:1.1M , bw:4.4GB )
"CompleteIO"(27027), index:0x201, filter:"AIR_READ" Period(iops:1.1M , bw:4.4GB , count:[{"4096": 3624199}]), Cumulation(iops:1.1M , bw:4.4GB )
( ) +Group:COMPLETE
( )[O]++Node:LAT_COMPLETE("latency")
""(0), index:0x0, filter:"AIR_0~AIR_1" Period(avg:92ns , median:67ns , max:512ns, sample:100 ), Cumulation(avg:152ns, median:99ns , max:5.8us, sample:500 )
( )[.]..Node:Q_COMPLETION
( ) +Group:SUBMIT
( )[O]++Node:CNT_TEST_EVENT("count")
"CompleteIO"(27027), index:0x0, filter:"AIR_COMPLETE" Period(count:3.6M ), Cumulation(count:14.4M )
"CompleteIO"(27027), index:0x1, filter:"AIR_COMPLETE" Period(count:-3.6M ), Cumulation(count:-14.3M )
( )[O]++Node:LAT_SUBMIT("latency")
""(0), index:0x0, filter:"AIR_0~AIR_1" Period(avg:236ns, median:136ns, max:940ns, sample:100 ), Cumulation(avg:331ns, median:195ns, max:7.1us, sample:482 )
( )[O]++Node:Q_SUBMISSION("queue")
"ProcessIO"(27025), index:0x0, filter:"AIR_BASE" Period(avg:251.4 , max:256.0 ), Cumulation(avg:250.8 , max:256.0 )
"ProcessIO"(27026), index:0x1, filter:"AIR_BASE" Period(avg:251.9 , max:256.0 ), Cumulation(avg:251.5 , max:256.0 )
In TUI, user can control AIR CLI in real-time.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent