Calculation and regulation of packet rates

Calculation and regulation of packet rates

Packet rates specified for Maximum Bandwidth or Guaranteed Bandwidth are:

rate = amount / time

where rate is expressed in kilobits per second (Kb/s).

Burst size at any given instant cannot exceed the amount configured in Maximum Bandwidth. Packets in excess are dropped. Packets deduct from the amount of bandwidth available to subsequent packets and available bandwidth regenerates at a fixed rate. As a result, bandwidth available to a given packet may be less than the configured rate, down to a minimum of 0 Kb/s.

Rate calculation and behavior can alternatively be described using the token bucket metaphor, where:

• A traffic flow has an associated bucket, which represents burst size bounds, and is the size of your configured bandwidth limit.
• The bucket receives tokens, which represent available bandwidth, at the fixed configured rate.
• As time passes, tokens are added to the bucket, up to the capacity of the bucket; excess tokens are discarded.
• When a packet arrives, the packet must deduct bandwidth tokens from the bucket equal to its packet size in order to egress.
• Packets cannot egress if there are insufficient tokens to pay for its egress; these nonconforming packets are dropped.

Bursts are not redistributed over a longer interval, so bursts are propagated rather than smoothed, although their peak size is limited.

Maximum burst size is the capacity of the bucket (the configured bandwidth limit); actual size varies by the current number of tokens in the bucket, which may be less than bucket capacity, due to deductions from previous packets and the fixed rate at which tokens accumulate. A depleted bucket refills at the rate of your configured bandwidth limit. Bursts cannot borrow tokens from other time intervals. This behavior is illustrated in the graph below.

Bursts and bandwidth limits over time

By limiting traffic peaks and token regeneration in this way, the available bandwidth at any given moment may be less than bucket capacity, but your limit on the total amount per time interval is ensured. Total bandwidth use during each interval of 1 second is at most the integral of your configured rate.

You may observe that external clients, such as FTP or BitTorrent clients, initially report rates between Maximum Bandwidth and twice that of Maximum Bandwidth, depending on the size of their initial burst. This is notably so when a connection is initiated following a period of no network activity.The apparent discrepancy in rates is caused by a difference in perspective when delimiting time intervals. A burst from the client may initially consume all tokens in the bucket, and before the end of 1 second, as the bucket regenerates, be allowed to consume almost another bucket’s worth of bandwidth. From the perspective of the client, this constitutes one time interval. From the perspective of the FortiGate unit, however, the bucket cannot accumulate tokens while full; therefore, the time interval for token regeneration begins after the initial burst, and does not contain the burst. These different points of reference result in an initial discrepancy equal to the size of the burst — the client’s rate contains it, but the FortiGate unit’s rate does not. If the connection is sustained to its limit and time progresses over an increasing number of intervals, however, this discrepancy decreases in importance relative to the bandwidth total, and the client’s reported rate will eventually approach that of the FortiGate unit’s configured rate limit.

For example, your Maximum Bandwidth might be 50 Kb/s and there has been no network activity for one or more seconds. The bucket is full. A burst from an FTP client immediately consumes 50 Kb. Because the bucket completely regenerates over 1 second, by the time almost another 1 second has elapsed from the initial burst, traffic can consume another 49.999 Kb, for a total of 99.999 Kb between the two points in time. From the vantage point of an external FTP client regulated by this bandwidth limit, it therefore initially appears that the bandwidth limit is 99.999 Kb/s, almost twice the configured limit of 50 Kb/s. However, bucket capacity only regenerates at your configured rate of 50 Kb/s, and so the connection can only consume a maximum of 50 Kb during each second thereafter. The result is that as bandwidth consumption is averaged over an increasing number of time intervals, each of which are limited to 50 Kb/s, the effects of the first interval’s doubled bandwidth size diminishes proportionately, and the client’s reported rate eventually approaches your configured rate limit. The effects are shown in the table below.

Effects of a 50 Kb/s limit on client reported rates

Guaranteed Bandwidth can also be described using a token bucket metaphor. However, because this feature attempts to achieve or exceed a rate rather than limit it, the FortiGate unit does not discard non-conforming packets, as it does for Maximum Bandwidth; instead, when the flow does not achieve the rate, the FortiGate unit increases the packets’ priority queue, in an effort to increase the rate.

Guaranteed and maximum bandwidth rates apply to the bidirectional total for all sessions controlled by the security policy. For example, an FTP connection may entail two separate connections for the data and control portion of the session; some packets may be reply traffic rather than initiating traffic. All packets for both connections are counted when calculating the packet rate for comparison with the guaranteed and maximum bandwidth rate.

Important considerations

By implementing QoS, you trade some performance and/or stability from traffic X by discarding packets or introducing latency in order to improve performance and stability of traffic Y. The best traffic shaping configuration for your network will balance the needs of each traffic flow by considering not only the needs of your particular organization, but also the resiliency and other characteristics of each particular service.

For example, you may find that web browsing traffic is both more resistant to interruptions or latency and less business critical than UDP or VoIP traffic, and so you might implement less restrictive QoS measures on UDP or VoIP traffic than on HTTP traffic.

An appropriate QoS configuration will also take into account the physical limits of your network devices, and the interactions of the aforementioned QoS mechanisms, described in Bandwidth guarantee, limit, and priority interactions on page 2468.

You may choose to configure QoS differently based upon the hardware limits of your network and FortiGate unit. Traffic shaping may be less beneficial in extremely high-volume situations where traffic exceeds a network interface’s or your FortiGate model’s overall physical capacity. A FortiGate unit must have enough resources, such as memory and processing power, to process all traffic it receives, and to process it at the required rate; if it does not have this capacity, then dropped packets and increased latency are likely to occur. For example, if the total amount of memory available for queuing on a physical interface is frequently exceeded by your network’s typical packet rates, frames and packets must be dropped. In such a situation, you might choose to implement QoS using a higher model FortiGate unit, or to configure an incoming bandwidth limit on each interface.

Incorrect traffic shaping configurations can actually further degrade certain network flows, because excessive discarding of packets or increased latency beyond points that can be gracefully handled by that protocol can create additional overhead at upper layers of the network, which may be attempting to recover from these errors. For example, a configuration might be too restrictive on the bandwidth accepted by an interface, and may therefore drop too many packets, resulting in the inability to complete or maintain a SIP call.

To optimize traffic shaping performance, first ensure that the network interface’s Ethernet statistics are clean of errors, collisions, or buffer overruns. To check the interface, enter the following diagnose command to see the traffic statistics:

diagnose hardware deviceinfo nic <port_name>

If these are not clean, adjust FortiGate unit and settings of routers or other network devices that are connected to the FortiGate unit. For more information, see Troubleshooting traffic shaping on page 2509.

Once Ethernet statistics are clean, you may want to use only some of the available FortiGate QoS techniques, or configure them differently, based upon the nature of FortiGate QoS mechanisms described in Bandwidth guarantee, limit, and priority interactions on page 2468.

Configuration considerations include:

• For maximum bandwidth limits, ensuring that bandwidth limits at the source interface and/or the security policy are not too low, which can cause the FortiGate unit to discard an excessive number of packets.
• For prioritization, considering the ratios of how packets are distributed between available queues, and which queue is used by which types of services. If you assign most packets to the same priority queue, it negates the effects of configuring prioritization. If you assign many high bandwidth services to high priority queues, lower priority queues may be starved for bandwidth and experience increased or indefinite latency. For example, you may want to prioritize a latency-sensitive service such as SIP over a bandwidth-intensive service such as FTP. Consider also that bandwidth guarantees can affect the queue distribution, assigning packets to queue 0 instead of their typical queue in high-volume situations.
• You may or may not want to guarantee bandwidth, because it causes the FortiGate unit to assign packets to queue
  • if the guaranteed packet rate is not currently being met. Comparing queuing behavior for lower-bandwidth and higher-bandwidth situations, this would mean that effects of prioritization only become visible as traffic volumes rise and exceed their guarantees. Because of this, you might want only some services to use bandwidth guarantees, to avoid the possibility that in high-volume situations all traffic uses the same queue, thereby negating the effects of configuring prioritization.

• For prioritization, configure prioritization for all through traffic. You may want to configure prioritization by either ToS-based priority or security policy priority, but not both. This simplifies analysis and troubleshooting.

Traffic subject to both security policy and ToS-based priorities will use a combined priority from both of those parts of the configuration, while traffic subject to only one of the prioritization methods will use only that priority. If you configure both methods, or if you configure either method for only a subset of your traffic, packets for which a combined priority applies will frequently receive a lower priority queue than packets for which you have only configured one priority method, or for which you have not configured prioritization.

For example, if both ToS-based priority and security policy priority both dictate that a packet should receive a “medium” priority, in the absence of bandwidth guarantees, a packet will use queue 3, while if only ToS-based priority had been configured, the packet would have used queue 1, and if only security policy-based priority had been configured, the packet would have used queue 2. If no prioritization had been configured at all, the packet would have used queue 0.