Queueing in the Linux Network Stack
Starvation and Latency
Despite its necessity and benefits, the queue between the IP stack and the hardware introduces two problems: starvation and latency.
If the NIC driver wakes to pull packets off of the queue for transmission and the queue is empty, the hardware will miss a transmission opportunity, thereby reducing the throughput of the system. This is referred to as starvation. Note that an empty queue when the system does not have anything to transmit is not starvation—this is normal. The complication associated with avoiding starvation is that the IP stack that is filling the queue and the hardware driver draining the queue run asynchronously. Worse, the duration between fill or drain events varies with the load on the system and external conditions, such as the network interface's physical medium. For example, on a busy system, the IP stack will get fewer opportunities to add packets to the queue, which increases the chances that the hardware will drain the queue before more packets are queued. For this reason, it is advantageous to have a very large queue to reduce the probability of starvation and ensure high throughput.
Although a large queue is necessary for a busy system to maintain high throughput, it has the downside of allowing for the introduction of a large amount of latency.
Figure 4 shows a driver queue that is almost full with TCP segments for a single high-bandwidth, bulk traffic flow (blue). Queued last is a packet from a VoIP or gaming flow (yellow). Interactive applications like VoIP or gaming typically emit small packets at fixed intervals that are latency-sensitive, while a high-bandwidth data transfer generates a higher packet rate and larger packets. This higher packet rate can fill the queue between interactive packets, causing the transmission of the interactive packet to be delayed.
Figure 4. Interactive Packet (Yellow) behind Bulk Flow Packets (Blue)
To illustrate this behaviour further, consider a scenario based on the following assumptions:
A network interface that is capable of transmitting at 5 Mbit/sec or 5,000,000 bits/sec.
Each packet from the bulk flow is 1,500 bytes or 12,000 bits.
Each packet from the interactive flow is 500 bytes.
The depth of the queue is 128 descriptors.
There are 127 bulk data packets and one interactive packet queued last.
Given the above assumptions, the time required to drain the 127 bulk packets and create a transmission opportunity for the interactive packet is (127 * 12,000) / 5,000,000 = 0.304 seconds (304 milliseconds for those who think of latency in terms of ping results). This amount of latency is well beyond what is acceptable for interactive applications, and this does not even represent the complete round-trip time—it is only the time required to transmit the packets queued before the interactive one. As described earlier, the size of the packets in the driver queue can be larger than 1,500 bytes, if TSO, UFO or GSO are enabled. This makes the latency problem correspondingly worse.
Large latencies introduced by over-sized, unmanaged queues is known as Bufferbloat (http://en.wikipedia.org/wiki/Bufferbloat). For a more detailed explanation of this phenomenon, see the Resources for this article.
As the above discussion illustrates, choosing the correct size for the driver queue is a Goldilocks problem—it can't be too small, or throughput suffers; it can't be too big, or latency suffers.
Byte Queue Limits (BQL)
Byte Queue Limits (BQL) is a new feature in recent Linux kernels (> 3.3.0) that attempts to solve the problem of driver queue sizing automatically. This is accomplished by adding a layer that enables and disables queueing to the driver queue based on calculating the minimum queue size required to avoid starvation under the current system conditions. Recall from earlier that the smaller the amount of queued data, the lower the maximum latency experienced by queued packets.
It is key to understand that the actual size of the driver queue is not changed by BQL. Rather, BQL calculates a limit of how much data (in bytes) can be queued at the current time. Any bytes over this limit must be held or dropped by the layers above the driver queue.
A real-world example may help provide a sense of how much BQL affects the amount of data that can be queued. On one of the author's servers, the driver queue size defaults to 256 descriptors. Since the Ethernet MTU is 1,500 bytes, this means up to 256 * 1,500 = 384,000 bytes can be queued to the driver queue (TSO, GSO and so forth are disabled, or this would be much higher). However, the limit value calculated by BQL is 3,012 bytes. As you can see, BQL greatly constrains the amount of data that can be queued.
BQL reduces network latency by limiting the amount of data in the driver queue to the minimum required to avoid starvation. It also has the important side effect of moving the point where most packets are queued from the driver queue, which is a simple FIFO, to the queueing discipline (QDisc) layer, which is capable of implementing much more complicated queueing strategies.
Dan Siemon is a longtime Linux user and former network admin who now spends most of his time doing business stuff.
Editorial Advisory Panel
Thank you to our 2014 Editorial Advisors!
- Jeff Parent
- Brad Baillio
- Nick Baronian
- Steve Case
- Chadalavada Kalyana
- Caleb Cullen
- Keir Davis
- Michael Eager
- Nick Faltys
- Dennis Frey
- Philip Jacob
- Jay Kruizenga
- Steve Marquez
- Dave McAllister
- Craig Oda
- Mike Roberts
- Chris Stark
- Patrick Swartz
- David Lynch
- Alicia Gibb
- Thomas Quinlan
- Carson McDonald
- Kristen Shoemaker
- Charnell Luchich
- James Walker
- Victor Gregorio
- Hari Boukis
- Brian Conner
- David Lane