Skip to main content

Provider Labels and Job Affinities/Uses

Labels​

All providers are tagged with labels, either defined by the provider, or generated through a benchmark. Labels are key-value elements.

Standard labels are:

LabelsDescriptionExample
osOperating system of the cluster.linux
archCPU architecture.amd64
cpuCPU reported by lscpu.amd-epyc-7302-16-core-processor
cpu.microarchCPU micro-architecture.zen2
gpuGPU of the cluster.nvidia-geforce-rtx-3090
compute.gflopsGiga Floating Points Operations Per Second (GFLOPS) of the whole cluster, based on the benchmark HPL-MXP-AI.96280.00 (96.280 TFLOPS)
network.download.bw.mbpsDownload bandwidth with the Internet in Mbps.1119.06
network.upload.bw.mbpsUpload bandwidth with the Internet in Mbps.1170.66
network.p2p.bw.mbpsPoint-to-Point bandwidth between two nodes in Mbps.40241.29
network.p2p.latency.usPoint-to-Point latency between two nodes in µs.202.53
network.all-to-all.latency.usBidirectional broadcast latency in µs.143.83
storage.scratch.read.bw.mibps$STORAGE_PATH read bandwidth in MiB/s.8927.18
storage.scratch.read.iops$STORAGE_PATH read I/O per seconds.4493.06
storage.scratch.write.bw.mibps$STORAGE_PATH write bandwidth in MiB/s.3862.60
storage.scratch.write.iops$STORAGE_PATH write I/O per seconds.1967.37
storage.shared-world-tmp.read.bw.mibps$DEEPSQUARE_SHARED_WORLD_TMP read bandwidth in MiB/s.8050.95
storage.shared-world-tmp.read.iops$DEEPSQUARE_SHARED_WORLD_TMP read I/O per seconds.4047.36
storage.shared-world-tmp.write.bw.mibps$DEEPSQUARE_SHARED_WORLD_TMP write bandwidth in MiB/s.4460.33
storage.shared-world-tmp.write.iops$DEEPSQUARE_SHARED_WORLD_TMP write I/O per seconds.2243.52
storage.shared-tmp.read.bw.mibps$DEEPSQUARE_SHARED_TMP read bandwidth in MiB/s.8619.40
storage.shared-tmp.read.iops$DEEPSQUARE_SHARED_TMP read I/O per seconds.4335.14
storage.shared-tmp.write.bw.mibps$DEEPSQUARE_SHARED_TMP write bandwidth in MiB/s.3920.92
storage.shared-tmp.write.iops$DEEPSQUARE_SHARED_TMP write I/O per seconds.1971.63
storage.disk-world-tmp.read.bw.mibps$DEEPSQUARE_DISK_WORLD_TMP read bandwidth in MiB/s.108934.02
storage.disk-world-tmp.read.iops$DEEPSQUARE_DISK_WORLD_TMP read I/O per seconds.54486.40
storage.disk-world-tmp.write.bw.mibps$DEEPSQUARE_DISK_WORLD_TMP write bandwidth in MiB/s.940.15
storage.disk-world-tmp.write.iops$DEEPSQUARE_DISK_WORLD_TMP write I/O per seconds.470.09
storage.disk-tmp.read.bw.mibps$DEEPSQUARE_DISK_TMP read bandwidth in MiB/s.104976.84
storage.disk-tmp.read.iops$DEEPSQUARE_DISK_TMP read I/O per seconds.52503.92
storage.disk-tmp.write.bw.mibps$DEEPSQUARE_DISK_TMP write bandwidth in MiB/s.786.46
storage.disk-tmp.write.iops$DEEPSQUARE_DISK_TMP write I/O per seconds.393.24

Affinities and Use flags when submitting jobs​

A job can filter the clusters by using affinities. An affinity is basically a rule. Affinities are a set of rules which limits the selection of clusters. A Use flag is a simplified notation of the affinity using the = operator.

An affinity is represented by a key, a value and an operator.

The key is used to select a label of a provider, the value is used for the comparison, and the operator is the type of rule.

There are 6 supported operators: < (less than), <= (less than or equal), = or == (equal), > (greater than), >= (greater than or equal), in (includes).

On the CLI, the in operator is wrapped with :. Example: gpu:in:nvidia.

For example, there were two clusters:

  • Cluster A, labelled:
    • arch=amd64
    • compute.gflops=1000
  • Cluster B, labelled:
    • arch=arm64
    • compute.gflops=10000

When submitting a job, it is possible to filter clusters:

  • "Only if the CPU architecture is arm64", rule is arch=arm64, which is also equivalent to the Use flag arch=arm64.
  • "Only if the GFLOPS is greater than 1000", rule is compute.gflops>1000.
  • "Only if the GPU is NVIDIA", rule is gpu:in:nvidia.

Next steps​

Now that you've learned about DeepSquare's filtering system, you can finish off by reading about the credits and SQUARE token used in our system to reward suppliers and add value to computing resources.