redis cluster百万QPS的挑战

最近在做redis cluster性能测试过程中，发现当集群吞吐量到达一定程度后（4台12core的redis服务器，80wQPS左右），集群整体性能不能线性增长。也就是说，通过加机器不能提升集群的整体吞吐。以下是详细记录了一下这个case的排查并最终解决的过程。

先来看一个压测图：

上图中每一条线代表一个起压端进程的压测QPS（一台起压机上开4个起压端），可以看到随着起压机的增多，每个起压机的QPS却在下滑。而此时redis的CPU、内存、网卡都还没有跑满。另外尝试增多redis服务器，发现增多一倍服务器后，每台服务器上每个redis实例的CPU大幅降低。总的QPS却基本没有变化。

由于起压机是无状态的，只是简单的mock随机请求到redis cluster，并且请求的路由策略会在起压机内部调jedis的过程中做好，直接请求到目标redis实例，并且在起压过程中没有槽迁移导致的ASK/MOVED动作。所以排除了起压机的影响，只能怀疑是redis cluster服务端导致的问题。

一开始怀疑是服务端连接数过多的问题，看了一下，服务端连接数在1w以下，每个客户端与每个redis实例的连接数也就在几十个长连接，远没有到达系统瓶颈。

难道是CPU的问题？但是top看每个核的CPU也就打到60%左右，而且如果是CPU到瓶颈了，通过增加redis服务器应该可以线性伸缩才对，但并没有。

再仔细观察top信息，发现每个核的CPU都有一定的si开销，也就是软中断，中断开销在10%以上：


怀疑是软中断导致抢占了Redis的CPU资源，软中断耗在哪里了可以用以下命令看到：

watch -d -n 1 'cat /proc/softirqs'

发现中断开销主要集中在网卡中断上，也就是eth0-TxRx上面。解决思路是将网卡中断和Redis进程使用的CPU隔离开，这样避免网卡中断抢占CPU拖慢Redis服务。另外，由于Redis是单线程模型，在多核服务器上，可以将多个Redis实例绑定到不同的CPU上，使每个Redis进程独占一个CPU。

先将网卡的中断号绑定到特定CPU上：

# 如果开了irqbalance服务，需要先停止服务，否则后续的绑定将无效：
service irqbalance stop

# 将网卡中断号绑定到CPU0-CPU2上：
echo "1" > /proc/irq/78/smp_affinity
echo "1" > /proc/irq/79/smp_affinity
echo "2" > /proc/irq/80/smp_affinity
echo "2" > /proc/irq/81/smp_affinity
echo "2" > /proc/irq/82/smp_affinity
echo "4" > /proc/irq/83/smp_affinity
echo "4" > /proc/irq/84/smp_affinity
echo "4" > /proc/irq/85/smp_affinity

再将Redis master实例绑定到CPU3-CPU10上。这里没有绑定Redis slave，是因为观察到slave占用的CPU极低，每个slave占用大概一个核的5%以下。当然如果CPU核数够多的话也可以考虑绑slave到其它多余的核。

# 绑定master的pid到CPU3-CPU10上：
taskset -cp 3 [pid1]
taskset -cp 4 [pid2]
taskset -cp 5 [pid3]
...

绑定好后再压，看top会发现，前三个CPU主要耗在si上（网卡中断），后8个CPU耗在us/sy上（redis master服务）：


至此，解决了单机CPU的问题，但是集群不能线性增长的问题还是解释不通。因为即使单台redis服务器的软中断导致单机性能受影响，在加Redis机器时总吞吐也不应该上不去呀。所以，开始怀疑redis cluster的通讯总线导致线性增长受阻。

每个redis cluster实例都会开一个集群通讯总线端口，并且redis cluster会用gossip协议，每隔一秒钟将集群 一部分节点的信息发送给某个实例的总线端口。利用这种传播方式，可以将集群节点的变更以无中心的方式传播到整个集群。这里关键是总线传播的数据量，数据量的大小取决于PING消息携带的gossip数组的长度和大小，这个长度就是具体要捎带多少个其它节点的信息：

/** cluster.c的clusterSendPing函数 **/

/* Send a PING or PONG packet to the specified node, making sure to add enough
 * gossip informations. */
void clusterSendPing(clusterLink *link, int type) {
    unsigned char *buf;
    clusterMsg *hdr;
    int gossipcount = 0; /* Number of gossip sections added so far. */
    int wanted; /* Number of gossip sections we want to append if possible. */
    int totlen; /* Total packet length. */
    /* freshnodes is the max number of nodes we can hope to append at all:
     * nodes available minus two (ourself and the node we are sending the
     * message to). However practically there may be less valid nodes since
     * nodes in handshake state, disconnected, are not considered. */
    int freshnodes = dictSize(server.cluster->nodes)-2;

    /* How many gossip sections we want to add? 1/10 of the number of nodes
     * and anyway at least 3. Why 1/10?
     *
     * If we have N masters, with N/10 entries, and we consider that in
     * node_timeout we exchange with each other node at least 4 packets
     * (we ping in the worst case in node_timeout/2 time, and we also
     * receive two pings from the host), we have a total of 8 packets
     * in the node_timeout*2 falure reports validity time. So we have
     * that, for a single PFAIL node, we can expect to receive the following
     * number of failure reports (in the specified window of time):
     *
     * PROB * GOSSIP_ENTRIES_PER_PACKET * TOTAL_PACKETS:
     *
     * PROB = probability of being featured in a single gossip entry,
     *        which is 1 / NUM_OF_NODES.
     * ENTRIES = 10.
     * TOTAL_PACKETS = 2 * 4 * NUM_OF_MASTERS.
     *
     * If we assume we have just masters (so num of nodes and num of masters
     * is the same), with 1/10 we always get over the majority, and specifically
     * 80% of the number of nodes, to account for many masters failing at the
     * same time.
     *
     * Since we have non-voting slaves that lower the probability of an entry
     * to feature our node, we set the number of entires per packet as
     * 10% of the total nodes we have. */
    wanted = floor(dictSize(server.cluster->nodes)/10);
    if (wanted < 3) wanted = 3;
    if (wanted > freshnodes) wanted = freshnodes;

    /* Compute the maxium totlen to allocate our buffer. We'll fix the totlen
     * later according to the number of gossip sections we really were able
     * to put inside the packet. */
    totlen = sizeof(clusterMsg)-sizeof(union clusterMsgData);
    totlen += (sizeof(clusterMsgDataGossip)*wanted);
    /* Note: clusterBuildMessageHdr() expects the buffer to be always at least
     * sizeof(clusterMsg) or more. */
    if (totlen < (int)sizeof(clusterMsg)) totlen = sizeof(clusterMsg);
    buf = zcalloc(totlen);
    hdr = (clusterMsg*) buf;

    /* Populate the header. */
    if (link->node && type == CLUSTERMSG_TYPE_PING)
        link->node->ping_sent = mstime();
    clusterBuildMessageHdr(hdr,type);

    /* Populate the gossip fields */
    int maxiterations = wanted*3;
    while(freshnodes > 0 && gossipcount < wanted && maxiterations--) {
        dictEntry *de = dictGetRandomKey(server.cluster->nodes);
        clusterNode *this = dictGetVal(de);
        clusterMsgDataGossip *gossip;
        int j;

        /* Don't include this node: the whole packet header is about us
         * already, so we just gossip about other nodes. */
        if (this == myself) continue;

        /* Give a bias to FAIL/PFAIL nodes. */
        if (maxiterations > wanted*2 &&
            !(this->flags & (REDIS_NODE_PFAIL|REDIS_NODE_FAIL)))
            continue;

        /* In the gossip section don't include:
         * 1) Nodes in HANDSHAKE state.
         * 3) Nodes with the NOADDR flag set.
         * 4) Disconnected nodes if they don't have configured slots.
         */
        if (this->flags & (REDIS_NODE_HANDSHAKE|REDIS_NODE_NOADDR) ||
            (this->link == NULL && this->numslots == 0))
        {
            freshnodes--; /* Tecnically not correct, but saves CPU. */
            continue;
        }

        /* Check if we already added this node */
        for (j = 0; j < gossipcount; j++) {
            if (memcmp(hdr->data.ping.gossip[j].nodename,this->name,
                    REDIS_CLUSTER_NAMELEN) == 0) break;
        }
        if (j != gossipcount) continue;

        /* Add it */
        freshnodes--;
        gossip = &(hdr->data.ping.gossip[gossipcount]);
        memcpy(gossip->nodename,this->name,REDIS_CLUSTER_NAMELEN);
        gossip->ping_sent = htonl(this->ping_sent);
        gossip->pong_received = htonl(this->pong_received);
        memcpy(gossip->ip,this->ip,sizeof(this->ip));
        gossip->port = htons(this->port);
        gossip->flags = htons(this->flags);
        gossip->notused1 = 0;
        gossip->notused2 = 0;
        gossipcount++;
    }

    /* Ready to send... fix the totlen fiend and queue the message in the
     * output buffer. */
    totlen = sizeof(clusterMsg)-sizeof(union clusterMsgData);
    totlen += (sizeof(clusterMsgDataGossip)*gossipcount);
    hdr->count = htons(gossipcount);
    hdr->totlen = htonl(totlen);
    clusterSendMessage(link,buf,totlen);
    zfree(buf);
}

以上函数定义了发送PING包的逻辑，重点是wanted变量的计算方式，也就是每次PING包内要携带多少个其他节点信息发送。可以看到，wanted最小3个，最大不超过集群总实例数-2个，一般情况是 集群实例数的1/10个。

再来看看每个gossip包的大小：

typedef struct {
    /* REDIS_CLUSTER_NAMELEN是常量40 */
    char nodename[REDIS_CLUSTER_NAMELEN];
    uint32_t ping_sent;
    uint32_t pong_received;
    char ip[REDIS_IP_STR_LEN];  /* IP address last time it was seen */
    uint16_t port;              /* port last time it was seen */
    uint16_t flags;             /* node->flags copy */
    uint16_t notused1;          /* Some room for future improvements. */
    uint32_t notused2;
} clusterMsgDataGossip;

看到这里就明白了，总线传播的数据量wanted是一个变量，会随着集群实例数的增多而增多。当集群节点数不多时不会造成影响，但随着节点数变多，每次wanted要携带的节点信息也就随之变多，会导致对网卡的消耗越来越大。我压测时每台Redis服务器8主8从共16个节点，四台服务器共64个节点，每个节点每秒钟PING包都要携带6个gossip包的信息（64/10）。如果扩大到十台，就是每个节点每秒钟PING包要携带16个gossip包的信息（160/10）。

解决了网卡中断问题后，Redis实例的CPU不会再被中断抢走，QPS基本可以线性增长。在6台12coreCPU的Redis服务器上，总QPS可以到达200w/s以上，单机QPS在30w以上。

ps：
推荐一篇有关软中断的文章： http://huoding.com/2013/10/30/296