gflofst
commented
4 years ago This is a parallel file system design decision. DAOS, IIRC, does not have distinct metadata servers compared to something like Lustre that has a separate metadata management portion of the storage system. Yes, the data is a bit messy. We have a student working to clean up the data and seek clarification where it is needed.
I'm having some trouble understanding the detail configuration information for the top 5 ranked IO500 any node list. Specifically the information for the md and ds sections in the complete data csv file.
For example, 1)the MD_store information for the DAOS #1 and #3 both show 0s for MD_nodes and all blank for _storage_devices, _vol_mem, _stor_interfac, _network and _os. But it shows 60 and 24 for ds_stor_devices and _stor_type is Optane PMEM for ds_stor_type. The Optane PMEM was used for MD storage and and md_storaga_interface should probably have been DIMM, DDR-T or PMEM-App direct. And there's no information about any (NVMe SSD) storage used by these two systems. 2) the md_volatile memory was listed as 384 for the #4 submission, not 384GB, and the md_stor_tape was listed as Intel Apache Pass 128GB DIMM and NVMe for the #2 ranked submission