Printer FriendlyEmail Article Link

Spirent TestCenter: How to get accurate Latency and Jitter measurement with STCv?

Environment/Versions
  • Spirent TestCenter.
Answer
In order to obtain the most accurate latency and jitter measurements, the following steps are recommended. Details are provided in the subsections below.
  1. Utilize a stratum one source
  2. Enable NTP Burst
  3. Allow NTP to sync over a longer period
  4. Use recommended deployments
  5. Follow hypervisor configuration recommendations
Enable NTP Burst
STCv Command to enable NTP Burst – ntp x.x.x.x burst
  • Provides 100 +/- Microsecond range with accuracy, when using a private NTP server synced to a stratum one clock.
Recommended Deployments for Latency MeasurementsRecommended Deployments for Latency Measurements
In order to eliminate clock differences, deploy a single VM with multiple test ports. If that is not feasible, the next best option is to deploy VMs on the same hypervisor.
  • Traffic between two STCv ports on the same STCv VM.
  • Traffic between two STCv VMs on the same Hypervisor.
  • Traffic between two STCv VMs across two sites.
  • Two NTP Servers running independently (Not Recommended).

Hypervisor Configuration
To increase the accuracy of latency measurements, you should:
  1. Use dedicated CPU cores for STCv VM
  2. Configure the hypervisor to isolate the STCv cpu cores with CPU pinning
    • Not having pinned vCPUs can lead to issues with ‘CPU Steal Time’
  3. Do not allow hypervisor interrupt processing on the STCv cores
Negative Latency
Given the nature of NTP, it is possible to get negative latency numbers even after following the recommendations. A negative latency is seen if the following ae true –
  1. The clocks on one of the VMs is ahead of the clock on another VM. This clock difference is inherent in the nature of NTP.
  2. The network latency is smaller than the clock difference.
For example, the picture below shows an example of negative latency
Latency in one direction: 538,409.92 us
Latency in the other direction: - 540,741.32 us (Negative latency)
In such cases, taking an average of the two latencies cancels out the clock difference and gives an accurate measure of average network latency:
Avg Latency = 538,409.92 + -540,741.32 = 2,331.4 /2 = 1,165.7 us.
 
Taking the Average
Why does taking the average cancel out the clock difference? Let’s look at the example below. Let’s say that the actual network latency between the two VMs is ‘L’, and the clock difference between the two VMs is ‘X’. In other words, when the clock on VM1 shows time t0, the clock on VM2 will show time t0+X.

Let’s say that Packet A is traveling from VM1 to VM2. It departs VM1 at time t0. Note that VM2’s clock would show a time of t0+X at that moment. The packet would arrive at VM2 after a period ‘L’ and would get an arrival timestamp of ‘t0+X+L’. Similarly, Packet B, which is traveling from VM2 to VM1, will have departure and arrival timestamps of ‘t0+X’ and ‘t0+L’ respectively.
Latency, which is determined by subtracting departure timestamp from the arrival timestamp, would be:
 
VM1 to VM2 latency (Packet A) = (t0+X+L) – (t0) = L+X
VM2 to VM1 Latency (Packet B) = (t0+L) – (t0+X) = L-X
 
Therefore, if VM2’s clock is ahead of VM1, the latency measured in VM1 -> VM2 direction will be higher and the latency measured in VM2 -> VM1 direction will be lower. If the clock difference X is larger than the network latency L, then we would see negative numbers for latency.

We can, however, determine the true latency L by averaging the two unidirectional latency numbers.
Avg latency = ((L+X) + (L-X))/2 = L
Attachments

Product : STC Virtual,Spirent TestCenter