SC25: The IO500 Lists Show Interesting Results as Usual

Two times a year, at SC in November and at ISC in June, the Top500 is updated with its IO500 companion.The last iteration unveiled a few days ago confirmed several things and presented as well new elements.

Benchmarks exist to compare systems and solutions, from public or private domains, and publish results for specific workloads. We can say that we see two types of benchmarks, the first to beat records and very often the config is exclusive and is not really affordable and the second one, more common, stressing systems on users’ workloads, data and environments. Then price rationalization is introduced in addition of absolute numbers.

SC25 updated its traditional lists in 5 categories: Research, 10 Node Research, Production, 10 Node Production and Full. We pay attention to the first 50 entries for each list to check new comers when it is possible.

• Research: 2 new entries with 1st LRZ at position 6 based on Lenovo systems coupled with DAOS and 2nd UW-Madison NetLab Research Group with MadRacks Testbed with GluonFS
• 10 Node Research: 6 new entries with obviously the previous one, 1st, LRZ, at position 10, 2nd, Barcelona Supercomputing Center at position 25 with Marenostrum 5 and GekkoFS, 3rd, UW-Madison NetLab Research Group, at position 27, 4th and 5th HPE at position 32 and 35 with HPE-DL320-SS400 and HPE-DL360-NDR both running DAOS, and 6th IBM Tucson at position 46 with Saguaro Storage Scale System 6000
• Production: 2 new systems with again LRZ landing at position 2 with SuperMUC-NG-Phase2-EC running DAOS and IBM at position 14 with Blue Vela Storage Scale System 6000 – shared coupled with IBM Storage Scale
• 10 Node Production: 10 new systems with interesting presence, 1st LRZ, the previous listed one at position 2, 2nd, Core42 at position 3 with Maximus-01 running DAOS, 3rd, JD Explore Academy at position 4 with JoyBuilder AI Development Service Platform running JPFS, 4th, SoftBank Corp at position 8 with CHIE-4 coupled with DDN ExaScaler, 5th, University of Florida at position 10 with HiPerGator running DDN ExaScaler, 6th, Simon Fraser University at position 15 with FIR running also DDN ExaScaler, 7th, Samsung Electronics at position 18 with Samsung AE-25 Cluster coupled with Hammerspace, 8th, Purdue University at position 20 again with DDN ExaScaler, 9th, TU Dresden ZIH at position 23 with Alpha Centauri coupled with Quobyte and 10th, Research Computing at Rochester Institute of Technology at position 30 with sporc running Ceph
• Full: The full list of course list several times LRZ with Lenovo and DAOS at position 12, 17, 25 and 34. Again we limit our filter to the first 50 entries

Clearly asymmetric storage model dominates with several parallel file system design. DDN ExaScaler is well represented but also Weka, IBM Storage Scale, Quobyte, BeeGFS, Vdura or Lustre even Ceph, some exotic solutions, but also lots of DAOS and we see some interesting appearances from NFS vendors such as Hammerspace validating their pNFS approach but also Vast Data and Qumulo. We notice that Hammerspace jumped immediately to position 18 as the 1st standard NFS model – here pNFS – in the 10 Node Production list and we expect other pNFS implementations to join the group for next iterations. To summarize, we can say that DAOS appears to be the winner across all these lists followed by DDN ExaScaler and then a group with a few entries.