I remember when...
The VAX 11/780 was the hot machine at One MIPS. It's time to retire me.
Cray has won a $50m-plus contract to provide 1 exabyte of ClusterStor storage for Oak Ridge National Laboratory's (ORNL) Frontier exascale supercomputer in the United States. Frontier is a $600m-plus Cray-AMD exascale system, rated at up to 1.5 exaFLOPS, due to be delivered in 2021 with acceptance in 2022. It will be a follow- …
The Vax 11/780 was *never* competing at this sort of level. The Vax was a significant bump in the performance that a small firm could afford, but it was never a supercomputer
The Cray-1 came out just before the Vax 11/780. The Cray could do about 130 MFlops.
(Conveniently that is almost exactly one tenth the performance of Raspberry Pi model B - but of course a Pi doesn't come with a seat.)
That's assuming they are using the formal definition of EB as decimal quantities. If they are using the technically incorrect but widely adopted practice of using EB, PB and TB as binary quantities then 1 EB = 1024 PB etc. Once you get up to PBs and EBs it starts to make a material difference...
There was a court case in the US some years ago regarding consumer storage that established that when the packaging said "MB", buyers had a reasonable expectation that this meant 1024 GBs (and so on down).
Am I missing something? You are SURE it's only ONE EXABYTE (One Thousand PetaBytes?)
That isn't actually all that large .... The LARGEST actual datacentre is the Intelligence Community Comprehensive National Cybersecurity Initiative Data Center located at a U.S. Military base called Camp Williams near Bluffdale, Utah which is which is at the Lat/Long Degrees:Minutes:Seconds GPS coordinates of:
or when using Decimal Degrees GPS coordinates of:
I actually remember from my contacts at IBM and Hitachi, that it was ZettaBytes of disk space that were ordered from multiple vendors!
That is MANY THOUSANDS of ExaBytes!
And based upon the PHYSICAL SIZE of the NSA data centre buildings and the size of 3.5 inch Enterprise-class drives, ZettaBytes is PROBABLY the more correct number of bytes of actual data stored by the NSA for both SHORT-TERM USE (6 months and less) and for longer-term use (Years and Decades) multi-path and multi-target correlation-oriented data mining!
I've got MULTIPLE ExaBytes of Tape Archive data in a room barely the size of a typical Mom and Pop Retail store! And that is just at my normal daily rendering facility! Yeah! That ONE mere ExaByte can't be right! I have that right next door to my office! For Oak Ridge National Lab, which is a MAJOR FEDERAL-level RESEARCH LABORATORY in the USA, a mere ExaByte is just toy storage! That must have been a misprint! It should be in the Zettabytes range for disk cluster storage......
OR ... maybe I'm just jaded (Yada Yada Yada! or Meh!) cuz of all the fancy custom-built compute and storage gear I've got here and that one exabyte is just a Meh! data storage size unit!
They may have that much in -total- storage, they aren't partitioning it as one big virtual volume like what ORNL is doing.
The point of this storage is to keep a single data set while it is being digested by the supercomputer, they may even have it set as a single contiguous mount point / partition. The intelligence services are going to break their storage into small chunks, one for each data set they are working with.
Yeah a 1EB Lustre file system, how long do they expect that to last before it eats itself. While Lustre has fsck tools they basically take forever in my experience to the point that even on a 150TB file system they are unusable, and Lustre would basically develop corruption about once a year for no apparent reason.
This would be, what, the third, forth time they've done that? Seems like every time they buy a supercomputer company, they just kinda sit on it, let it rot, and then mercy kill the product line. Then they'll buy a new HPC company, then do the same.
Supercomputer Shmoopercomputer. These are almost a joke because there is only so much parallelism in the universe of modelling. To quote an old colleague, I can vectorize the code (problem) to 99%, I can parallelize the code to 99%, but I cannot do both at the same time so all those processors and GPUs reflect prurient interests for the buyers and sellers, and serious challenges for the users.
Conclusion, these so-called supercomputers are essentially big, fancy time sharing systems where the application could be equally served by smaller systems, and the workload by a number of smaller systems. As for the huge file system(s)? Where is my data, how did it get there, and how do I get it back - need another few thousand processors for that :D ?