Scaling OpenShift Origin HAProxy Router

Baseline Performance

The OpenShift Origin router is the ingress point for all external traffic destined for OpenShift Origin services.

On an public cloud instance of size 4 vCPU/16GB RAM, a single HAProxy router is able to handle between 7000-32000 HTTP keep-alive requests depending on encryption, page size, and the number of connections used. For example, when using TLS edge or re-encryption terminations with large page sizes and a high numbers of connections, expect to see results in the lower range. With HTTP keep-alive, a single HAProxy router is capable of saturating 1 Gbit NIC at page sizes as small as 8 kB.

The table below shows HTTP keep-alive performance on such a public cloud instance with a single HAProxy and 100 routes:

Encryption Page size HTTP(s) requests per second

none

1kB

15435

none

4kB

11947

edge

1kB

7467

edge

4kB

7678

passthrough

1kB

25789

passthrough

4kB

17876

re-encrypt

1kB

7611

re-encrypt

4kB

7395

When running on bare metal with modern processors, you can expect roughly twice the performance of the public cloud instance above. This overhead is introduced by the virtualization layer in place on public clouds and holds mostly true for private cloud-based virtualization as well. The following table is a guide on how many applications to use behind the router:

Number of applications Application type

5-10

static file/web server or caching proxy

100-1000

applications generating dynamic content

In general, HAProxy can saturate about 5-1000 applications, depending on the technology in use. The number will typically be lower for applications serving only static content.

Router sharding should be used to serve more routes towards applications and help horizontally scale the routing tier.

Performance Optimizations

Setting the Maximum Number of Connections

One of the most important tunable parameters for HAProxy scalability is the maxconn parameter, which sets the maximum per-process number of concurrent connections to a given number. Adjust this parameter by editing the ROUTER_MAX_CONNECTIONS environment variable in the OpenShift Origin HAProxy router’s deployment configuration file.

CPU and Interrupt Affinity

In OpenShift Origin, the HAProxy router runs as a single process. The OpenShift Origin HAProxy router typically performs better on a system with fewer but high frequency cores, rather than on an symmetric multiprocessing (SMP) system with a high number of lower frequency cores.

Pinning the HAProxy process to one CPU core and the network interrupts to another CPU core tends to increase network performance. Having processes and interrupts on the same non-uniform memory access (NUMA) node helps avoid memory accesses by ensuring a shared L3 cache. However, this level of control is generally not possible on a public cloud environment. On bare metal hosts, irqbalance automatically handles peripheral component interconnect (PCI) locality and NUMA affinity for interrupt request lines (IRQs). On a cloud environment, this level of information is generally not provided to the operating system.

CPU pinning is performed either by taskset or by using HAProxy’s cpu-map parameter. This directive takes two arguments: the process ID and the CPU core ID. For example, to pin HAProxy process 1 onto CPU core 0, add the following line to the global section of HAProxy’s configuration file:

    cpu-map 1 0

To modify the HAProxy configuration file, refer to Deploying a Customized HAProxy Router.

Impacts of Buffer Increases

The OpenShift Origin HAProxy router request buffer configuration limits the size of headers in incoming requests and responses from applications. The HAProxy parameter tune.bufsize can be increased to allow processing of larger headers and to allow applications with very large cookies to work, such as those accepted by load balancers provided by many public cloud providers. However, this affects the total memory use, especially when large numbers of connections are open. With very large numbers of open connections, the memory usage will be nearly proportionate to the increase of this tunable parameter.