New DISCO switch telemetry dataset

Since June 2016, M-Lab has collected high resolution switch telemetry for each M-Lab server and site uplink.

Originally designed to detect switch discards from server traffic microbursts, we now support the DIScard COllection (a.k.a. DISCO) dataset as a standard M-Lab BigQuery table: measurement-lab.base_tables.switch

DISCO Metrics

To see the complete list of DISCO metrics, run the following query:

#standardSQL
SELECT
    metric
FROM
    `measurement-lab.utilization.switch`
GROUP BY
    metric
ORDER BY
    metric ASC

DISCO metric names use a common structure: switch.<type>.<source>.<direction>.

DISCO collects seven standard metric types: packets sent by broadcast, packets discarded by the switch, packet errors, packets sent by multicast, octet counts (e.g. bytes), pause frames (for Ethernet flow control), and unicast packets.

DISCO collects metrics for two sources: machine local and switch uplink. The “local” metrics report values from the switch port attached to the machine collecting the switch data, e.g. hostname = "mlab1.sea02.measurement-lab.org". So there will be “local” metrics for each machine at a site. The “uplink” metrics report values from the switch port attached to the uplink router, outside M-Lab’s administrative domain. Each machine collects these so there is some built-in redunancy.

Finally, a DISCO metric indicates whether the value is for data received (rx) by the port or transferred (tx) from the port. Note that the data direction is relative the the switch port. So, for example, a “download” test will show up on the switch.unicast.local.rx followed by the switch.unicast.uplink.tx.

All sample.values are counts over the 10 second window starting at the given sample.timestamp.

Example Queries

The switch schema takes advantage of record arrays. So, queries that use the sample.value or sample.timestamp columns must handle the array records appropriately.

Daily Average Download Rates

The following query calculates the per-day, per-host average data output rate from sites in the DFW metro area. The native sample.value is a 10-second counter delta. So, to calculate the per-second rate, we must divide by 10. And, the unit of switch.octets.uplink.tx metric is bytes, so we must multiply by 8 to convert to bits.

#standardSQL
SELECT
  TIMESTAMP_TRUNC(sample.timestamp, DAY) AS day,
  hostname,
  metric,
  8 * ROUND(AVG(sample.value) / 1e6, 1) / 10 AS mbps
FROM
  `measurement-lab.utilization.switch`,
  UNNEST(sample) AS sample
WHERE
  metric = "switch.octets.uplink.tx"
  AND hostname LIKE "%mlab1.dfw%"
GROUP BY
  day,
  hostname,
  metric
ORDER BY
  day

Daily 99th Percentile Discard Rate

The following query calculates the 99th percentile packet discard rate per day. Because we want a per-second rate, we must divide sample.value by 10. And, we use APPROX_QUANTILES in part to exclude extreme values caused by erroneous samples in the raw data.

#standardSQL
SELECT
  TIMESTAMP_TRUNC(sample.timestamp, DAY) AS day,
  hostname,
  metric,
  APPROX_QUANTILES(sample.value / 10, 100)[OFFSET(99)] AS pps
FROM
  `measurement-lab.utilization.switch`,
  UNNEST(sample) AS sample
WHERE
  metric = "switch.discards.uplink.tx"
  AND hostname LIKE "%mlab1.dfw%"
GROUP BY
  day,
  hostname,
  metric
ORDER BY
  day

Percentage of Download Traffic Per-machine

Of the data leaving a site, this query roughly estimates the percentage attributed to each machine. The percentage may change, for example, when one machine is offline and traffic is diverted to another.

For simplicity, this query looks at a single site over time. For most of the history, mlab-ns was using the “first online server” policy, so mlab1 sees about 100% of traffic during that time. However after June 1st, 2018, the mlab-ns server selection algorithm was changed to “random online server”, and the ratios approach 1/3.

#standardSQL
SELECT
  day,
  100 * mlab1_total/uplink_total as mlab1,
  100 * mlab2_total/uplink_total as mlab2,
  100 * mlab3_total/uplink_total as mlab3
FROM (
SELECT
  TIMESTAMP_TRUNC(sample.timestamp, DAY) AS day,
  SUM(IF(metric = "switch.octets.local.rx" AND hostname like "%mlab1%", sample.value, 0)) as mlab1_total,
  SUM(IF(metric = "switch.octets.local.rx" AND hostname like "%mlab2%", sample.value, 0)) as mlab2_total,
  SUM(IF(metric = "switch.octets.local.rx" AND hostname like "%mlab3%", sample.value, 0)) as mlab3_total,
  SUM(IF(metric = "switch.octets.uplink.tx" AND hostname like "%mlab1%", sample.value, 0)) as uplink_total
FROM
  `measurement-lab.utilization.switch`,
  UNNEST(sample) AS sample
WHERE
  (metric = "switch.octets.uplink.tx" OR metric = "switch.octets.local.rx")
  AND hostname LIKE "%yyz%"
GROUP BY
  day
)
WHERE
  uplink_total > 0
ORDER BY
  day