Skip to content

Latest commit

 

History

History
103 lines (79 loc) · 3.51 KB

runtime_metrics.md

File metadata and controls

103 lines (79 loc) · 3.51 KB
Raw

Runtime on different hardware configurations

Recommendations

The recommended compute setup for DeepConsensus is to shard each SMRT Cell into at least 500 shards, each of which can run on a 16-CPU machine (or smaller). We find that having more than 16 CPUs available for each shard does not significantly improve runtime. Additionally, running with --batch_size=1024 --batch_zmws=100 is a reasonable setup for minimizing memory requirements while still achieving the best runtime. If you find that this exhausts the memory on your machines though, reducing both, e.g. to --batch_size=512 --batch_zmws=20 would be a good thing to try.

See the yield metrics page for total runtimes on full SMRT Cells of different sequencing runs.

Testing runtime on different hardware configurations

We processed 10,000 ZMWs with ccs and actc. After filtering, we were left with 3,577 ZMWs, which we used to profile runtimes using deepconsensus run across several hardware configurations. These estimates only reflect the runtime required to perform inference, after subread data has been preprocessed.

DeepConsensus runtime profiling

Runtime, max memory usage, batch_size, and batch_zmws are shown across different machine types. We observe that batch_zmws has a large impact on memory usage. Our current implementation requires setting batch_zmws and batch_size carefully to achieve optimal performance. We are working to improve performance further and allow for more predictable runtimes based on DeepConsensus settings.

We only show GPU runtimes for n1-standard-16. Note that we observe job failures when using GPU with larger batch_size (>=4096) and with larger values of batch_zmws (500, 1000).

Runtime Test Configurations

n2-standard-64: 64vCPUs (Cascade Lake)

This command shows what machine we tested on:

gcloud compute instances create "${USER}-n2-64" \
  --scopes "compute-rw,storage-full,cloud-platform" \
  --image-family "ubuntu-2004-lts" \
  --image-project "ubuntu-os-cloud" \
  --machine-type "n2-standard-64" \
  --boot-disk-size "200" \
  --zone "us-west1-b"

n2-standard-16: 16vCPUs (Cascade Lake)

This command shows what machine we tested on:

gcloud compute instances create "${USER}-n2-16" \
  --scopes "compute-rw,storage-full,cloud-platform" \
  --image-family "ubuntu-2004-lts" \
  --image-project "ubuntu-os-cloud" \
  --machine-type "n2-standard-16" \
  --boot-disk-size "200" \
  --zone "us-west1-b"

n1-standard-16, 16vCPUs

This command shows what machine we tested on:

gcloud compute instances create "${USER}-gpu" \
  --scopes "compute-rw,storage-full,cloud-platform" \
  --maintenance-policy "TERMINATE" \
  --image-family "ubuntu-2004-lts" \
  --image-project "ubuntu-os-cloud" \
  --machine-type "n1-standard-16" \
  --boot-disk-size "200" \
  --zone "us-west1-b"

n1-standard-16 + P100 GPU, 16vCPUs (SkyLake)

This command shows what machine we tested on:

gcloud compute instances create "${USER}-gpu" \
  --scopes "compute-rw,storage-full,cloud-platform" \
  --maintenance-policy "TERMINATE" \
  --accelerator=type=nvidia-tesla-p100,count=1 \
  --image-family "ubuntu-2004-lts" \
  --image-project "ubuntu-os-cloud" \
  --machine-type "n1-standard-16" \
  --boot-disk-size "200" \
  --zone "us-west1-b" \
  --min-cpu-platform "Intel Skylake"

Test Dataset

The 10k ZMW test dataset is available on Google Storage:

gs://brain-genomics-public/research/deepconsensus/quickstart/v0.3/n10000.subreads.bam