agent是指标采集模块,仅关注linux本身的监控指标,主要负责:

  • 定期进行指标采集,然后通过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的指标结果发送给Transfer;

整体架构:

1. 指标采集

代码入口:

func main() {
    ......
    cron.Collect()
    ......
}

//modules/agent/cron/collector.go
func Collect() {
    .....
    for _, v := range funcs.Mappers {
        go collect(int64(v.Interval), v.Fs)
    }
}

其中funcs.Mappers是采集函数的集合,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.go
var Mappers []FuncsAndInterval
func 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,将采集的指标汇集起来,发送给transfer:

// modules/agent/cron/collector.go
func 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,只要上报成功就退出,不再尝试其他的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长连接,由SingleConnRpcClient来维护:

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")完成最终的方法调用;

  • 首先确保有1个TCP rpc长连接到Transfer;
  • 调用rpc.Client.Call()方法进行实际的rpc调用;
  • rpc调用放在goroutine内执行,外面使用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的类型语义相同:

  • GAUGE表示有实际意义的立即数,比如内存用量;
  • COUNTER表示递增的计数,比如cpu使用时间的计数;

agent用cpu在user模式下的使用计数差值 / 总模式的计数差值,得到cpu.user.percent(GAUGE类型),具体来看一下代码;

agent使用1个goroutine来定期采集/proc/stat下各种cpu的使用计数:

//modules/agent/cron/collector.go
func InitDataHistory() {
    for {
        funcs.UpdateCpuStat()        
        time.Sleep(g.COLLECT_INTERVAL)
    }
}

由于要计算统计时间间隔内的差值,故保存了2份数据,上一次统计和本次统计的:

//modules/agent/funcs/cpustat.go
const (
    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指标数据的来源,是读取/proc/stats文件:

//github.com/toolkits/nux/cpustat.go
func 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
}

最后看下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.go
func 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方法(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查询自己要执行的plugin名称及版本信息,通过调用hbs的RPC方法(Agent.MinePlugins)实现:

// modules/agent/cron/plugin.go
func 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出特定格式的指标信息,agent读取该执行结果,发送到transfer;比如:

[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查询要执行的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.go
func 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,然后读取cmd命令的输出结果,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的执行有个风险,由于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是对应系统线程,较多的M被创建并阻塞,会导致系统的性能下降。

参考:
1.Open-falcon docs: https://book.open-falcon.org/...


a朋
63 声望38 粉丝

引用和评论

0 条评论