Let's say I have the following setup:
client (browser)-> load balancer (nginx) -> 10 nodejs servers
I'm using websocket (socket.io) for bidirectional communication between client and server.
My question is: when a client sends a message via socket, does load balancer simply redirect the socket request to another server like a normal http request, or something more complicated is going on, like making the socket remain open on the load balancer machine for the whole duration of the websocket connection between client and server.
Because if the above is true, it means that load balancer machine has the SUM of ALL its child servers' TCP connection opened, which means load balancer has to be quite a huge machine on it own, and it's not "entirely" distributing all the load.. and eventually I have to worry about "running out of sockets" on the load balancer machine?
Can anyone clarify this whole concept for me? Thanks!
For websockets, yes, the load balancer will almost certainly maintain a TCP connection in from the client and out to the server for every connection. The load balancer, however, has much less work to do than the server machines actually handling the clients -- it doesn't have to "think" about the protocol or generate or interpret any payload, it just has to copy data from one pipe to another and tear down connections when they close.
With an event-driven model and potentially using the Linux kernel's splice(2) system call, the "copied" data can be shuffled between the connections all in kernel space, for very efficient operation.
A well-designed load balancer is as likely to run up against the IPv4 limit of ~64K address/port pairs for a given IP address as any other resource, like CPU or memory.
Even with normal web traffic, where the balancer is understanding and making routing decisions based on request and response bodies, the app servers still typically have far more work to do than the balancer.
Anecdotally, some of the smallest machines in my infrastructure are the load balancers... and they still tend to have the smallest workloads as evidenced by memory, CPU, and disk access.
Related
Every few months when thinking through a personal project that involves sockets I find myself having the question of "How would you properly load balance sockets on a dynamic horizontally scaling WebSocket server?"
I understand the theory behind horizontally scaling the WebSockets and using pub/sub models to get data to the right server that holds the socket connection for a specific user. I think I understand ways to effectively identify the server with the fewest current socket connections that I would want to route a new socket connection too. What I don't understand is how to effectively route new socket connections to the server you've picked with low socket count.
I don't imagine this answer would be tied to a specific server implementation, but rather could be applied to most servers. I could easily see myself implementing this with vert.x, node.js, or even perfect.
First off, you need to define the bounds of the problem you're asking about. If you're truly talking about dynamic horizontal scaling where you spin up and down servers based on total load, then that's an even more involved problem than just figuring out where to route the latest incoming new socket connection.
To solve that problem, you have to have a way of "moving" a socket from one host to another so you can clear connections from a host that you want to spin down (I'm assuming here that true dynamic scaling goes both up and down). The usual way I've seen that done is by engaging a cooperating client where you tell the client to reconnect and when it reconnects it is load balanced onto a different server so you can clear off the one you wanted to spin down. If your client has auto-reconnect logic already (like socket.io does), you can just have the server close the connection and the client will automatically re-connect.
As for load balancing the incoming client connections, you have to decide what load metric you want to use. Ultimately, you need a score for each server process that tells you how "busy" you think it is so you can put new connections on the least busy server. A rudimentary score would just be number of current connections. If you have large numbers of connections per server process (tens of thousands) and there's no particular reason in your app that some might be lots more busy than others, then the law of large numbers probably averages out the load so you could get away with just how many connections each server has. If the use of connections is not that fair or even, then you may have to also factor in some sort of time moving average of the CPU load along with the total number of connections.
If you're going to load balance across multiple physical servers, then you will need a load balancer or proxy service that everyone connects to initially and that proxy can look at the metrics for all currently running servers in the pool and assign the connection to the one with the most lowest current score. That can either be done with a proxy scheme or (more scalable) via a redirect so the proxy gets out of the way after the initial assignment.
You could then also have a process that regularly examines your load score (however you decided to calculate it) on all the servers in the cluster and decides when to spin a new server up or when to spin one down or when things are too far out of balance on a given server and that server needs to be told to kick several connections off, forcing them to rebalance.
What I don't understand is how to effectively route new socket connections to the server you've picked with low socket count.
As described above, you either use a proxy scheme or a redirect scheme. At a slightly higher cost at connection time, I favor the redirect scheme because it's more scalable when running and creates fewer points of failure for an existing connection. All clients connect to your incoming connection gateway server which is responsible for knowing the current load score for each of the servers in the farm and based on that, it assigns an incoming connection to the host with the lowest score and this new connection is then redirected to reconnect to one of the specific servers in your farm.
I have also seen load balancing done purely by a custom DNS implementation. Client requests IP address for farm.somedomain.com and that custom DNS server gives them the IP address of the host it wants them assigned to. Each client that looks up the IP address for farm.somedomain.com may get a different IP address. You spin hosts up or down by adding or removing them from the custom DNS server and it is that custom DNS server that has to contain the logic for knowing the load balancing logic and the current load scores of all the running hosts.
Route the websocket requests to a load balancer that makes the decision about where to send the connections.
As an example, HAProxy has a leastconn method for long connections that picks the least recently used server with the lowest connection count.
The HAProxy backend server weightings can also be modified by external inputs, #jfriend00 detailed the technicalities of weighting in their answer.
I found this project that might be useful:
https://github.com/apundir/wsbalancer
A snippet from the description:
Websocket balancer is a stateful reverse proxy for websockets. It distributes incoming websockets across multiple available backends. In addition to load balancing, the balancer also takes care of transparently switching from one backend to another in case of mid session abnormal failure.
During this failover, the remote client connection is retained as-is thus remote client do not even see this failover. Every attempt is made to ensure none of the message is dropped during this failover.
Regarding your question : that new connection will be routed by the load balancer if configured to do so.
As #Matt mentioned, for example with HAProxy using the leastconn option.
My client makes two http requests to my cloud service which has two replicas.
According to documentation (1) and since connection is kept alive, I'd expect the two requests to go to the same replica.
However, I see each request goes to a different replica. For performance reasons, this is undesirable.
What is causing the distribution?
How do I debug load balancer?
(1) https://azure.microsoft.com/en-us/documentation/articles/load-balancer-distribution-mode/
The default distribution will be 5-tuple (SourceIP, Destination IP, source Port, Destination Port, Protocol). It means that each new connection initiated by a client may land to a different server
If you use sourceIP, then the stickiness will be based on the client IP address
If you need application based stickiness (such as cookie based affinity, then you may look at https://azure.microsoft.com/en-us/documentation/services/application-gateway/
Yves
I want to add a load balancer infront of my nodejs websockets server. The plan is to add another node on another physical machine and have a load balancer in front. The load balancer will also be on its own physical machine.
The requirement is that several 1000s of simultaneous connections could be handled and I'm a bit worried about bouncys upper limitations.
I like the consistency of using bouncy since it is a node module, but at the same time it seems like nginx could handle more socket connections or be a bit more stable.
Anyone who has experience with bouncy or nginx as load balancer and could give me some advices?
Thanks!
nginx is pretty good for mass connections, check these answer.
https://stackoverflow.com/a/16289251/2325522
there you can see how to use Nginx as load balacer.
The only problem that you can have is the mass band-width needed to serve 1000's of simultaneous connections.
Example:
5000 clients * 0.25Mb/request (a little one)
=
1250mb (1.25Gb outgoing band-width)
Hope these solve your doubts.
I have a clustered server with 4 nodes running Win server 2008 r2 with IIS 7.
Fail over kicks in when one of nodes fails but is there a way to have it round robin distribute incoming calls to different server?
This happens when incoming requests come from different client but our investigation shows that if there is one client that is making many requests, they all go to the same server.
I would like to the server to round robin request so that node 1 receives first request, node 2 receives second request and so on.
Each request could take a long time and having all requests go to the same node when I have 3 others idling is causing us perf issue. Thanks
NLB port rules have a couple of properties that control how requests are routed. The relevant properties seem to be:
Filtering mode - specifies whether a single host or multiple hosts in the cluster handle traffic for the given port
Affinity - controls how traffic is routed to hosts in the cluster
It is likely you need to set the Affinity value to none, which allows requests to be routed to multiple hosts within the cluster. The docs do not state whether round-robin or another algorithm is used for load balancing.
For more on Filtering Mode and Affinity: Network Load Balancing Manager Properties
How to: Edit a Network Load Balancing Port Rule
Round Robin Load Balancing will not distribute traffic coming from one destination. You will need to configure your load balancer to 'Least Connections'
Basically the NLB passes a new connection to the pool member or node that has the least number of active connections.
I don't have the faintest clue on how a software or hardware load balancer works. I guess the hardware load balancer is basically a switch and based on some algorithm decides which node to switch to for a incoming request. On the software load balancer front, I guess the software picks up a node and uses a reverse proxy connection to it. In such a scenario, 2-way SSL wont work as the load balancer cannot have the client's private key.
Again, I don't how a software load balancer works but as my application would need a load balancer and as the application uses 2-way SSL connection, I wanted to know how does a software load balancer take care of a 2-way SSL connection.
No, SSL works with a load balancer. They typically work at the TCP level, so the clients connect to the LB IP address, but it NATs the connections on to the real servers. The connection persists to the same real server for its lifetime, but if the same client makes another one, it can (and typically would) go to a different server.
For HTTPS this works fine, except that if you have a web server which supports SSL session caching, then the SSL session cache will be lost if the client comes back to a different server. In practice this is not a big problem. Of course HTTP keep-alive sessions aren't affected because they are a single TCP connection so they stay on the same realserver.
Generally speaking, a software load balancer will note that there is a new incoming connection request, assess the workload on the machines available, and allocate the new request to the most appropriate machine. When there is a session-based service, that connection will last for the duration of the session; rebalancing would only occur if a server went down, and would probably establish new connections in a newly balanced configuration.
So, as Jon implied, the SSL session would be established with a server, and would continue with that server until the session terminates.
If you want to route connections more dynamically, then it may be that the SSL session has to be terminated (decrypted) in front of the software that dynamically sends requests to different servers.
All these are possible - they are not necessarily efficient or implemented.
A software load balancer will distribute sessions evenly across multiple servers.
So, if a user hits your load balancer, it will send him to a specific server and that server will negotiate the SSL. The user will continually talk to this server until his session expires. At that point, he will hit the load balancer again.