Quic Timestamps For Measuring One-Way Delays

1.1. Terms and Definitions

The keywords "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶

2.1. Negotiation

The use of the timestamp frame extension is negotiated using a transport parameter:¶

• enable_timestamp (TBD)¶

The enable timestamp transport parameter is included if the endpoint wants to receive or accepts to send timestamp frames for this connection. This parameter is encoded as a variable integer as specified in section 16 of [QUIC-TRANSPORT]. It can take one of the following three values:¶

1. I would like to receive TIMESTAMP frames¶
2. I am able to generate TIMESTAMP frames¶
3. I am able to generate TIMESTAMP frames and I would like to receive them¶

Peers receiving another value SHOULD terminate the connection with a TRANSPORT PARAMETER error.¶

A peer that advertises its capability of sending TIMESTAMP frames using option values 2 or 3 MUST NOT send these frames if the other peer does not announce advertise its desire to receive them by sending the enable_timestamp TP with option 1 or 3. This condition is described as "successful sending negotiation" in Section 2.2.¶

Peers that receive TIMESTAMP frames when they have not advertised their desire to receive them MAY terminate the connection with a PROTOCOL VIOLATION error.¶

2.2. Sending TIMESTAMP frames

Following successful sending negotiation, a peer SHOULD add a timestamp frame to 1RTT packets carrying an ACK frame. This specification does not impose a placement of TIMESTAMP frames in the packet. They MAY be sent either before or after the ACK frame.¶

Implementations SHOULD NOT send more than one TIMESTAMP frame per packet, but they MAY send more than one in rare circumstances. When multiple TIMESTAMP frames are present in a packet, the receiver retains the frame indicating the largest timestamp.¶

Implementations MUST NOT send the TIMESTAMP frame in Initial, 0-RTT or Handshake packets, because there is a risk that the peer will receive such packets before the negotiation completes. This restriction may appear excessive because some Handshake packets are typically sent after the negotiation completes, but restricting TIMESTAMP frames to 1RTT packets is simpler and less error prone than allowing the TIMESTAMP frame in just a fraction of Handshake packets.¶

2.3. TIMESTAMP frame format

TIMESTAMP frames are identified by the frame type:¶

• TIMESTAMP (TBD)¶

TIMESTAMP frames carry a single parameter, the timestamp.¶

 0                   1                   2                   3
 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|                          Timestamp (i)                      ...
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+

The timestamp encodes the number of microseconds since the beginning of the epoch, as measured by the peer at the time at which the packet is sent. It is encoded using the exponent selected by the peer in the ack_delay_exponent. The exponent reduced timestamp is encoded as a variable length integer.¶

TIMESTAMP frames are not ack-eliciting. Their loss does not require retransmission.¶

For congestion control, TIMESTAMP frames are treated like ACK frames. Section 7 of [QUIC-RECOVERY] specifies that "packets containing only ACK frames do not count towards bytes in flight and are not congestion controlled". The same applies to packets containing only TIMESTAMP frames, or a combination of ACK frames and TIMESTAMP frames.¶

2.4. RTT Measurements

RTT measurements are performed as specified in Section 4 of [QUIC-RECOVERY], without reference to the Timestamp parameter of the Timestamped ACK frames.¶

2.5. Choice of Epoch

Each peer can chose its epoch as it sees fit, but it MUST remain constant for the duration of the connection, and the resulting timestamps MUST be positive integers. Plausible values for the epoch could be:¶

• the beginning of the connection, i.e., the time at which the first packet for that connection was sent or received.¶
• the time at which the first timestamp is sent.¶

Chosing values close to the beginning of the connection ensures that the timestamps value will be at most equal to the duration of the connection, which limits the amount of bytes required to encode the timestamps.¶

2.6. One-Way Delay Measurements

An endpoint generates a One Way Delay Sample on receiving a packet containing both a TIMESTAMP frame and an ACK frame that meets the following two conditions:¶

• the largest acknowledged packet number is newly acknowledged, and¶
• at least one of the newly acknowledged packets was ack-eliciting.¶

The One Way Delay sample, latest_1wd, is generated as the time elapsed since the largest acknowledged packet was sent, corrected for the difference between local time at the sending peer and connection time at the receiving peer, phase_shift.¶

latest_1wd = timestamp - send_time_of_largest_acked - phase_shift¶

By convention, the phase_shift is estimated upon reception of the first RTT sample, first_rtt. It is set to:¶

phase_shift = timestamp - send_time_of_largest_acked - latest_rtt/2¶

In that formula, we assume that the local time are measured in microseconds since the beginning of the connection. The formula does not depend on the choice of epoch by each peer, but simply of the hypothesis that delays on the data path and the return path are about equal.¶

We understand that clocks may drift over time, and that simply estimating a phase shift at the beginning of a connection may be too simplistic for long duration connections. Implementations MAY adopt different strategies to reestimate the phase shift at appropriate intervals. Specifying these strategies is beyond the scope of this document.¶

3.1. Management of Time

There are two known issues with deducing one way delays from RTT measurements: clock drift and undefined phase difference.¶

The phase difference problem is easy to understand. We start from a list of measurements associating the send time of packet number x (s[x]), the receive time of the acknowledgement of packet (a[x]), and the timestamp indicating when packet x was received by the peer (p[x]). The peer's timestamp are expressed in the peer's clock.¶

Suppose that we model the peer's clock as local time plus phase difference f, and that we model the rtt as the sum of two one way delays, up (u[x]) and down (d[x]). We get:¶

    u[x] = p[x] + f - s[x]

    d[x] = a[x] - p[x] - f

Just looking at the equation shows that the value of f cannot be determined from the a series of measurement (s[x], a[x], p[x]). You can just add constraints that all u[x] and d[x] are positive numbers, which gives a range of plausible values for f: max(s[x] - p[x]) < f < min(a[x]-p[x]). In case you wonder, you get similar formulations in a multipath scenario. The plausible range may narrow to the min rtt of the shortest path, but no further.¶

The phase difference uncertainty is not a big issue in practice, because control algorithms are much more interested in the variations of the delays than by their absolute values. Suppose we want to compare one way delays at measurement (x) and (y). We get:¶

    u[x] = p[x] + f - s[x]

    u[y] = p[y] + f - s[y]

    u[x] - u[y] = p[x] - p[y] - s[x] + s[y]

The phase difference does not affect the measurement of variations in the one way delay.¶

The clock drift is another matter. All the equations above assume that the local clock and the remote clock have the same frequency. This is an approximation. Clocks drift over time. Instead of just considering a stable phase difference, one should consider the sum of a phase difference and a time-varying drift component. Estimating drift is a complex problem. This was studied in detail in the development of the Network Time Protocol (NTP) [RFC5905]. In theory, implementations of Quic could copy the algorithms of NTP to build a model of the clocks used by the local node and the peer. That would be very complex.¶

Fortunately, implementations of Quic no not need to implement something as complex as NTP. Most time based algorithms are only interested in variations of delays over a short horizon. Clock drift happens at a slow pace, maybe 1 millisecond per minute. Time base congestion control algorithms already have to cope with the potential drift of the minimum RTT due to changing network conditions. They do that by periodically restarting the measurement of the minimum RTT after some delay, typically less than a minute. A simple implementation of one way delay measurements could follow the same approach, for example resetting the phase difference every 30 seconds or so.¶

4.1. Application to Hystart

Implementations of the Cubic congestion control [RFC8312] benefit from the Hybrid Slow Start algorithm [HyStart]. Hystart works by monitoring the transmission delay during the initial Slow Start phase, exiting slow start and moving to congestion avoidance when a delay increase is noticed, before usage of a too large congestion window causes many losses. Hystart generally improves performance of Cubic, but can cause performance degradation if spurious delay measurement cause an early exit.¶

An example of early exit was noticed when implementing Cubic congestion control in QUIC [CH09]. Tests over a wireless connection showed significant jitter in the RTT measurements. This was fixed by exiting only after 5 measurements showed a delay increase. Further investigation showed that this jitter was mostly happening on the "upload" path. Timestamps enable measurement of the delay variations independently on each path, and thus improved exit decisions.¶

4.2. Application to QUIC Multipath

Time Stamps are very useful in multipath environments, as mentioned in [MULTIPATH-QUIC]. In the absence of time stamps, it is very hard to estimate the contribution of each path to the end to end delay, and the specification mandates that acknowledgements be sent on the same path over which packets were received. If time stamps are available, experiments show that performance are improved if the acknowlegments are sent on the shortest available path, because the implementation can detect packet losses or congestion events faster.¶