It’s been a little over five years since Sherlock first entered production and ran its first jobs. A lot of time has passed, and with it, Sherlock grew ten-fold, its user base exploded and technology trends evolved significantly.

The old

The original Sherlock nodes are still running jobs, but they’re incrementally moving towards irrelevance as they lack the features, the performance and the power efficiency of the most recently added compute nodes.

This is why we’re starting the process of gradually retiring the original Sherlock nodes. This will mainly affect the Sherlock public tiers in the normal, gpu and bigmem partition, but also some of the earliest owner partitions (we’ve started contacting the owners of those nodes).

The most noticeable effect will be that the size of those partitions will start shrinking, and over time, those nodes and their 5-year old Ivy Bridge generation CPUs will progressively disappear.

To compensate for these retirements, we’re also introducing new nodes to the normal partition.

The new

The new nodes are based on the latest Intel Skylake microarchitecture and have the following specifications:

  • dual-socket Intel Xeon Gold 5118 (2x 12-core, 24 core per node)
  • 192GB of memory (RAM)
  • 200GB local SSD (for $L_SCRATCH)
  • EDR Infiniband (100Gbps) connectivity to the Sherlock fabric.

You’ll be able to request those nodes specifically by using job submission constraints and adding the following flag to your job submission options: -C "CPU_GEN:SKX"

To see the list of all the available node features and characteristics that can be requested in the normal partition, you can run:

$ sh_node_feat -p normal

And as always, please don’t hesitate to let us know if you have any question, by reaching out to [email protected].