agent是指标采集模块,仅关注linux自身的监控指标,次要负责:
- 定期进行指标采集,而后通过RPC上报给Transfer;
- 向hbs发送heartbeat,同时从hbs获取要监听的process、port和要执行的plugin信息;
- 定期执行plugin,将plugin的指标后果发送给Transfer;
整体架构:
1. 指标采集
代码入口:
func main() { ...... cron.Collect() ......}//modules/agent/cron/collector.gofunc Collect() { ..... for _, v := range funcs.Mappers { go collect(int64(v.Interval), v.Fs) }}
其中funcs.Mappers是采集函数的汇合,agent为每一类采集启动了一个goroutine,有几种分类就有几个goroutine:
//modules/agent/funcs/funcs.govar Mappers []FuncsAndIntervalfunc BuildMappers() { interval := g.Config().Transfer.Interval Mappers = []FuncsAndInterval{ { Fs: []func() []*model.MetricValue{ AgentMetrics, CpuMetrics, NetMetrics, KernelMetrics, LoadAvgMetrics, ...... }, Interval: interval, }, { Fs: []func() []*model.MetricValue{ DeviceMetrics, }, Interval: interval, }, ...... }}
具体的采集过程,执行每个采集函数,将采集的指标会集起来,发送给transfer:
// modules/agent/cron/collector.gofunc collect(sec int64, fns []func() []*model.MetricValue) { t := time.NewTicker(time.Second * time.Duration(sec)) defer t.Stop() for { <-t.C hostname, err := g.Hostname() if err != nil { continue } mvs := []*model.MetricValue{} ignoreMetrics := g.Config().IgnoreMetrics for _, fn := range fns { items := fn() for _, mv := range items { mvs = append(mvs, mv) ...... } } ...... g.SendToTransfer(mvs) }}
2. 指标上报Transfer
agent与Transfer之间是TCP RPC通道,agent配置了多个transfer的地址,上报时随机选一个地址,只有上报胜利就退出,不再尝试其余的transfer地址:
func SendMetrics(metrics []*model.MetricValue, resp *model.TransferResponse) { rand.Seed(time.Now().UnixNano()) for _, i := range rand.Perm(len(Config().Transfer.Addrs)) { addr := Config().Transfer.Addrs[i] c := getTransferClient(addr) if c == nil { c = initTransferClient(addr) } if updateMetrics(c, metrics, resp) { break } }}
agent与transfer之间维持一个TCP长连贯,由SingleConnRpcClient来保护:
type SingleConnRpcClient struct { sync.Mutex rpcClient *rpc.Client RpcServer string Timeout time.Duration}
SingleConnRpcClient包装了rpc.Client,调用rpc.Client.Call("Transfer.Update")实现最终的办法调用;
- 首先确保有1个TCP rpc长连贯到Transfer;
- 调用rpc.Client.Call()办法进行理论的rpc调用;
- rpc调用放在goroutine内执行,里面应用select进行超时判断;
func (this *SingleConnRpcClient) Call(method string, args interface{}, reply interface{}) error { this.Lock() defer this.Unlock() err := this.serverConn() if err != nil { return err } timeout := time.Duration(10 * time.Second) done := make(chan error, 1) go func() { err := this.rpcClient.Call(method, args, reply) done <- err }() select { case <-time.After(timeout): this.close() return errors.New(this.RpcServer + " rpc call timeout") case err := <-done: if err != nil { this.close() return err } } return nil}
3. 指标类型:GUAGE与COUNTER
GAUGE与COUNTER类型,与Prometheus的类型语义雷同:
- GAUGE示意有实际意义的立刻数,比方内存用量;
- COUNTER示意递增的计数,比方cpu应用工夫的计数;
agent用cpu在user模式下的应用计数差值 / 总模式的计数差值,失去cpu.user.percent(GAUGE类型),具体来看一下代码;
agent应用1个goroutine来定期采集/proc/stat下各种cpu的应用计数:
//modules/agent/cron/collector.gofunc InitDataHistory() { for { funcs.UpdateCpuStat() time.Sleep(g.COLLECT_INTERVAL) }}
因为要计算统计工夫距离内的差值,故保留了2份数据,上一次统计和本次统计的:
//modules/agent/funcs/cpustat.goconst ( historyCount int = 2)func UpdateCpuStat() error { ps, err := nux.CurrentProcStat() if err != nil { return err } psLock.Lock() defer psLock.Unlock() for i := historyCount - 1; i > 0; i-- { procStatHistory[i] = procStatHistory[i-1] } procStatHistory[0] = ps return nil}
cpu指标数据的起源,是读取/proc/stats文件:
//github.com/toolkits/nux/cpustat.gofunc CurrentProcStat() (*ProcStat, error) { f := Root() + "/proc/stat" bs, err := ioutil.ReadFile(f) if err != nil { return nil, err } ps := &ProcStat{Cpus: make([]*CpuUsage, NumCpu())} reader := bufio.NewReader(bytes.NewBuffer(bs)) for { line, err := file.ReadLine(reader) if err == io.EOF { err = nil break } else if err != nil { return ps, err } parseLine(line, ps) } return ps, nil}
最初看下cpu.user.percent指标如何计算的,其计算公式为:
(cpu.user2 - cpu.use1) / (cpu.total2 - cpu.total1)
//modules/agent/funcs/cpustat.gofunc CpuMetrics() []*model.MetricValue { user := GaugeValue("cpu.user", CpuUser()) return []*model.MetricValue{user, ....}}func CpuUser() float64 { psLock.RLock() defer psLock.RUnlock() dt := deltaTotal() if dt == 0 { return 0.0 } invQuotient := 100.00 / float64(dt) return float64(procStatHistory[0].Cpu.User-procStatHistory[1].Cpu.User) * invQuotient}func deltaTotal() uint64 { if procStatHistory[1] == nil { return 0 } return procStatHistory[0].Cpu.Total - procStatHistory[1].Cpu.Total}
4. agent与hbs
agent定期向hbs发送heartbeat,上报时调用hbs的RPC办法(Agent.ReportStatus)实现的:
func reportAgentStatus(interval time.Duration) { for { hostname, err := g.Hostname() req := model.AgentReportRequest{ Hostname: hostname, IP: g.IP(), AgentVersion: g.VersionMsg(), PluginVersion: g.GetCurrPluginVersion(), } var resp model.SimpleRpcResponse err = g.HbsClient.Call("Agent.ReportStatus", req, &resp) time.Sleep(interval) }}
agent还会定期向hbs查问本人要执行的plugin名称及版本信息,通过调用hbs的RPC办法(Agent.MinePlugins)实现:
// modules/agent/cron/plugin.gofunc syncMinePlugins() { duration := time.Duration(g.Config().Heartbeat.Interval) * time.Second for { time.Sleep(duration) hostname, err := g.Hostname() req := model.AgentHeartbeatRequest{ Hostname: hostname, } var resp model.AgentPluginsResponse err = g.HbsClient.Call("Agent.MinePlugins", req, &resp) ...... }}
5. 执行plugin
plugin即采集的插件或脚本,通常是shell或py,该脚本执行后通常会输入特定格局的指标信息,agent读取该执行后果,发送到transfer;比方:
[root@host01:/path/to/plugins/plugin/sys/ntp]#./600_ntp.py[{"endpoint": "host01", "tags": "", "timestamp": 1431349763, "metric": "sys.ntp.offset", "value": 0.73699999999999999, "counterType": "GAUGE", "step": 600}]
上一步讲到,agent向hbs查问要执行的plugin list,跟本地执行的plugin list进行比对,若有新的plugin则执行:
//modules/agent/plugins/plugins.gofunc AddNewPlugins(newPlugins map[string]*Plugin) { for fpath, newPlugin := range newPlugins { if _, ok := Plugins[fpath]; ok && newPlugin.MTime == Plugins[fpath].MTime { continue } Plugins[fpath] = newPlugin sch := NewPluginScheduler(newPlugin) PluginsWithScheduler[fpath] = sch sch.Schedule() }}
每个plugin是定期被执行的:
func (this *PluginScheduler) Schedule() { go func() { for { select { case <-this.Ticker.C: PluginRun(this.Plugin) case <-this.Quit: this.Ticker.Stop() return } } }()}
plugin的执行,实际上是执行cmd命令,而后读取cmd命令的输入后果,解析后发送给transfer:
func PluginRun(plugin *Plugin) { var cmd *exec.Cmd if args == "" { cmd = exec.Command(fpath) } else { arg_list := PluginArgsParse(args) cmd = exec.Command(fpath, arg_list...) } //执行cmd命令 err, isTimeout := sys.CmdRunWithTimeout(cmd, time.Duration(timeout)*time.Millisecond) // exec successfully data := stdout.Bytes() //发送到transfer var metrics []*model.MetricValue err = json.Unmarshal(data, &metrics) g.SendToTransfer(metrics)}
plugin的执行有个危险,因为goroutine执行cmd会进入SYSCALL阻塞,也就是其对应的G和M一起被阻塞,若阻塞较多的话,会有较多的M被阻塞,M是对应零碎线程,较多的M被创立并阻塞,会导致系统的性能降落。
参考:
1.Open-falcon docs: https://book.open-falcon.org/...