我有一个用Go编写的新TCP服务器,它连接了100多个客户端。每个客户端都会在需要集中查看的数据中进行流式传输,因为他们正在查看来自不同位置的空中波的无线电数据包,然后进行分析。代码工作但我看到很多争用和增加锁定周围的CPU,并考虑如何避免锁定(如果可能)或围绕它进行优化。
当TCP服务器为收到的每个数据包旋转GoRoutine时,addMessage
功能需要一定程度的同步。这些数据包也将在稍后的另一个函数中进行分析,该函数在地图上执行RLock()
。
cullMessages()
函数每秒被调用一次,它本身就会被捕获并且可能真的变慢,有时需要2-3秒才能运行,这会将问题复杂化为接下来的2-3次操作排队等待解锁并立即运行!
任何想法/想法都将不胜感激!
var dataMessagesMutex sync.RWMutex
var dataMessages map[string][]*trackingPacket_v1
// Function is called from each TCP client who need to share this data
func addMessage(trackingPacket *trackingPacket_v1) {
dataMessagesMutex.Lock()
dataMessages[trackingPacket.packetID] = append(dataMessages[trackingPacket.packetID], trackingPacket)
dataMessagesMutex.Unlock()
}
// Function called on a loop, need to delete based on age here
func cullMessages() {
cullTS := time.Now().Add(-time.Second * MODES_MAX_MESSAGE_AGE)
dataMessagesMutex.Lock()
defer dataMessagesMutex.Unlock()
for avr, data := range dataMessages {
sort.Sort(PacketSorter(data))
highestIndex := 0
for i, messages := range data {
if cullTS.Sub(messages.ProcessedTime) > 0 {
// Need to delete the message here
messages = nil
highestIndex = i
}
}
// Copy the new slice into the data variable
data = data[highestIndex+1:]
if len(data) == 0 {
// Empty Messages, delete
delete(dataMessages, avr)
}
}
}
更新: 添加了分析功能
func processCandidates() {
mlatMessagesMutex.RLock()
defer dataMessagesMutex.RUnlock()
for _, data := range dataMessages {
numberOfMessages := len(data)
for a := 0; a < numberOfMessages; a++ {
packetA := data[a]
applicablePackets := []*trackingPacket_v1{packetA}
for b := 0; b < numberOfMessages; b++ {
// Don't compare identical packets
if b == a {
continue
}
packetB := data[b]
// Only consider this packet if it's within an acceptable
// timestamp threshold
tsDelta := math.Abs(packetA.NormalisedTS - packetB.NormalisedTS)
if tsDelta < MAX_MESSAGE_TS_DIFF {
// Finally, we need to make sure that only one message per
// station is included in our batch
stationAlreadyRepresented := false
for i := 0; i < len(applicablePackets); i++ {
if applicablePackets[i].Sharecode == packetB.Sharecode {
stationAlreadyRepresented = true
}
}
if stationAlreadyRepresented == false {
applicablePackets = append(applicablePackets, packetB)
}
}
}
// Remove any stations which are deemed too close to one another
if len(applicablePackets) >= MIN_STATIONS_NEEDED {
applicablePackets = cullPackets(applicablePackets)
}
// Provided we still have enough packets....
if len(applicablePackets) >= MIN_STATIONS_NEEDED {
// Generate a hash for this batch...
hash := generateHashForPackets(applicablePackets)
batchIsUnique := true
for _, packet := range applicablePackets {
if packet.containsHash(hash) {
batchIsUnique = false
break
}
}
if batchIsUnique == true {
for _, packet := range applicablePackets {
packet.addHash(hash)
}
go sendOfDataForWork(applicablePackets)
}
}
}
}
}
答案 0 :(得分:1)
不是拥有一张大地图,而是为每个packetID设置一个goroutine。调度程序goroutine可以有map[string]chan *trackingPacket_v1
,并在适当的通道上发送传入的数据包。然后,该packetID的goroutine将数据包收集到本地切片中,并剔除它们并定期分析它们。
不知何故,您需要终止未在MODES_MAX_MESSAGE_AGE中收到数据包的goroutine。调度员goroutine可能会跟踪最近看到每个packetID的时间,并定期检查并检查那些太旧的数据包ID。然后它将关闭这些通道并将其从地图中删除。当分析goroutine发现其通道已关闭时,它将退出。