TCP Incast Problem on Data centers in Cloud Computing

In previous posts, we discussed cloud computing and its business models. Today, we are going to learn about effects of TCP Incast Problem on Data centers in Cloud Computing and solution for TCP Incast.

The Effects of TCP Incast Problem on Data centers in Cloud Computing

Cloud computing is a domain that requires big data centers to supports its applications and services. Companies like Amazon, E-Bay, Google, etc. use big data centers to cater the users with a wide variety of services like web search, e-commerce, and data storage.

In order to support huge amount of data traffic, data centers require high-capacity links (high burst tolerance), high throughput and low propagation delays.  TCP tends to be the most popularly used protocol over the internet. However, in data centers, TCP is not able to perform well due to the incast problem.
Incast occurs in cloud data centers due to many-to-one communication links. This happens when a server is connected to various nodes via a switch. When the server places a request for data, it is sent to all the nodes simultaneously, which in turn reply at the same time.

This causes a micro burst of all the data coming from the nodes towards a single server (many to one communication). Due to low queuing capacity at the server, the packets start dropping out as the data rate also increases drastically. This eventually leads to throughput collapse at the server side.

Solutions for TCP Incast

In order to reduce the TCP incast problem, switches having large buffers can be used, but it turns out to be a costly solution and also results in high latency as the buffers get deep.

Another way to cater TCP incast is to reduce TCP’s minimum RTO (Request Time Out), which will help TCP to enter in the retransmission mode quickly and deal with congestion as quickly as possible. However, RTO should not be decreased too much, as it will increase the retransmissions drastically and ultimately choke the bandwidth.

Data Center TCP (DCTCP) is a modified version of TCP, which is specifically designed to cater cloud computing applications and deal with high throughput. DCTCP is based on ECN – Explicit Congestion Notification, which takes congestion feedback and operates accordingly. A threshold value is marked and the number of packets exceeding the threshold value in the queue are marked. The window size is then decreased in accordance with the number of marked bits.

DCTCP works well because it is able to send packets before they are dropped, which results in high throughput. Moreover, DCTCP achieves low latency due to small buffer occupancies.

Apart from incast, TCP in data center environment also suffers from the problems of queue build up and buffer pressure. Due to big data flows, queues build up at the receiver and even short messages have to wait for a long time due to increased latency. Buffer pressure implies that the port bandwidth on a switch is adjusted according to the amount of data sent on it. So if one port on the switch has long flow and the other has short flow, so more bandwidth would be given to the long one and short one would automatically be affected.

Click here to read more.