How the CPU score is measured

Every offering, instance and host page shows a CPU score: how fast the host’s processor runs everyday work. It comes as two figures, the single thread score and, for a size, a total. Both are Gigahost’s own measurement of the host, not figures taken from a CPU database, and each host’s score comes only from that host’s own runs.

What the score measures

The score comes from a suite of eight small workloads that stand for what people run on a server:

• compression,
• matrix math,
• sorting,
• hash maps,
• JSON parsing,
• SQLite queries,
• a bytecode interpreter,
• an image filter.

Each workload runs on one core, so the single thread score tells how fast one core of the processor works and says nothing of how many cores the host has. The suite leaves out anything that depends on special hardware, such as encryption or SHA instructions, because one processor having a unit that another lacks says little about the work most instances do.

How often it runs

Each host runs the suite once a day, at a random time of its own, so the runs of the fleet do not all land at the same hour. The page shows the best score of the host’s last 7 runs. A run can only be slowed down by what else happens on the machine, such as a busy neighbour or a backup, and never made faster, so the best of several runs is the closest to what the processor can do.

A new version of the suite starts a host’s history again, because its scores are on another scale. Until a host has run the suite once, its pages say Not measured yet.

The scale

A run’s score is the geometric mean of its workloads’ ratios to a reference machine, scaled so that the reference scores 1696. The reference is an Intel Xeon E5-2683 v4, and 1696 is the single-thread rating PassMark gives that processor. That puts the numbers on a familiar scale: a score of 3400 means about twice the single-core speed of the reference. Every other processor is measured by Gigahost’s own suite, so its score is not a PassMark rating.

Hosts are never mixed

Runs of two hosts are never pooled. An instance and a host’s page show that host’s own score.

An offering can span several hosts of one provider with the same hardware profile, and a rental can be placed on any of them. Its page shows the average of their scores, rounded to a whole number: three hosts that score 3,800, 3,900 and 4,101 show 3,934. A host that has not run the suite yet leaves the average to the others, and an offering none of whose hosts has run it says Not measured yet. While a new version of the suite reaches the fleet, the average is taken over the hosts that report the newest version, so scores of two versions never meet.

The total figure

For a size, the page also shows a total: roughly what its vCPUs can do together. It is an approximation worked out from the single thread score, not a second measurement, because the real figure always depends on what the neighbours on the host are doing.

The total follows how a host spreads vCPUs over its processor:

• The first vCPU counts at the full single thread score.
• Further vCPUs land on cores of their own while the host has cores to spare. A small size runs on a quiet processor, so its first few vCPUs add nearly a whole score each. The more cores a size takes, the less each one adds, because every core runs slower when many of them work at once. When a size takes all of the host’s physical cores, each further core has added about two thirds of the score (0.67) on average.
• Only once the size has more vCPUs than the host has physical cores do the rest share a core with another thread, and each adds about a fifth of the score (0.18).

For a single thread score of 4000 on a host with 16 cores, 2 vCPUs give 7,934, 4 vCPUs 15,387, 16 vCPUs 4000 × (1 + 15 × 0.67) = 44,200, and 32 vCPUs 4000 × (1 + 15 × 0.67 + 16 × 0.18) = 55,720.

The whole-host figure was chosen to land within about 7 % of the multi-thread ratings PassMark publishes for recent desktop processors. Server processors lose more than that when every core works, so the page uses a whole-host share for each generation, from PassMark’s ratings of its processors: each further core of an Intel Xeon E5 adds about 0.42 to 0.50 on average (by generation), of an Intel Xeon Scalable about 0.47, of an AMD EPYC about 0.58. Small sizes on those processors still start at nearly a whole score per vCPU. For the largest EPYC parts, with 64 cores or more, the total is still on the high side.

The CPU tier’s sharing is left out of the total. It matters only while the host’s processor is fully loaded, and then the figure depends on the neighbours anyway.

What the score does not promise

The score describes the host’s processor. It is not a guarantee for an instance: an instance that shares the host’s cores with others runs as fast as the load on those cores allows, and the total is an approximation of the size, not a measurement of it.

The figure is what the host’s own agent reports after running the suite on its machine. Gigahost checks that each report is complete and within what a processor can reach, but it does not run the suite itself, so a score is only as good as the host behind it. A provider’s reliability record and reviews, which sit beside the score, say how far to trust the rest of what the host shows.