Thursday, August 25, 2016

Globally Synchronized Time Via Datacenter Networks

Presenter: Ki Suh Lee
Co-Authors: Han Wang, Vishal Shrivatsav, Hakim Weatherspoon

Synchronized Clocks are important for network and distributed systems since precise clocks help you with monitoring, coordination, updates, etc. Broadly, clock synchronization protocols work measure the offset or the time difference between two clocks and the precision or the maximum offset between any two clocks. The two servers exchange messages between them that contain the timestamps of the servers. Using this information and the round-trip time of the message, one is usually able to compute the offset between the two servers.

However, measuring the RTT is hard. There could be a skew, timestamping could be wrong or the network could induce delays in the messages sent between the two servers. PTP (Precision Time Protocol) tries to overcome this by hardware stamping on PTP-enabled switches, yet it doesn't provide bounded precision and requires PTP-enabled devices in the network. Approaches like NTP also suffer from much higher precision errors.

The authors propose DTP which uses the physical layer since clocks are already synchronized at that level between peers. Further, this approach doesn't impact higher layers and doesn't have network overhead either. This is able to achieve nanosecond level precisions since the physical layer clock is updated every few nanoseconds, resulting in high accuracies.

A DTP enabled device has a local counter that is updated at every clock tick or is adjusted upon the receipt of a message from a peer. The DTP protocol overwrites /E/ bocks in the PHY layer with the time at the physical layer and sends these messages to its peer. The message sending involves two phases namely an INIT and a BEACON. The two peers send an INIT message that is sent back to the sender as an ACK and using this, one can compute the one-way delay. The BEACON messages contains the local counter value of the sender and the receiver uses this information, its delay and its own local counter to update its clock. The error in this process is utmost 4 clock tick errors (2 from the one-way delay message and 2 from the beacon sending).
To synchronize this externally, a DTP Daemon could be used to poll these counter values and correlate them with UTC values. This DTP approach needs NIC and switch modifications like PTP, but gives better precision bounds than PTP.

To evaluate DTP, the authors used an FPGA development board along with few other modules and setup 12 identical servers in a tree topology. They measured offset between peers in this 3 level network. They also evaluated PTP on the same topology.  The evaluations showed that while the PTP offset is about 10s to 100s of nanosecond when idle, the offset is close to 100 microseconds when experiencing load. DTP offset, on the other hand, is always within 4 clock ticks and within a few nanoseconds. Even within a datacenter with 6 hops, the offset between peers is about 150 nanoseconds and end-to end application delay isn't more than 200 ns.

Q: Do you think the ability to hack into the physical layer is generally useful for other reasons? How easy was it to get into the physical layer?
A: It is not very hard to insert something into the physical layer, once you understand it and know what you are doing, you can modify the bits to do other things

Q: Does this translate into the wireless domain at all?
A: Wireless domain is a little bit different, there are collisions, wireless doesn't guarantee that something is being sent continuously.
Q: What if it was all point-to-point?
A: yes (no interference)

Q: You make some assumptions about the frequency. What if this was meant to run on a different link with different frequency?
A: Relationship between clock frequency. If you control the increments with the frequency carefully, it might still work.

Q: This needs to be standardized to be used. Any plans?
A: Looking forward to making this a standard for others to use.

Q: This mechanism keeps it internally synced, but could drift from a global counter. What are your thoughts on external synchronization?
A: Use a connected FPGA board that is synchronized via GPS and this acts as a global clock, but you could potentially synchronize all internal clocks to that Global clock.