3 ms·

But half the top supercomputers are a cluster of Linux x86 boxes, and the other half are Linux on some other architecture.

by comex 3y ago

datameta 3y ago
I think it's not accurate to call their architecture a cluster of machines. Either way I'm not referring to Linux as an OS - it is really the only choice for enterprise scale computing. Regardless of what architectures supercomputers use, their use cases are different from that of server/mainframe high performance compute. 8 cores with on-chip AI acceleration running at 5GHz sharing virtual cache, using ~256GB-1TB of of RAM at ~76-304Gbps transfer speed is something a major bank would use. It won't cut it for SoTA atmospheric modeling, molecular modeling, or something like rapid computational chemistry search space exploration. What it excels at is the ability to handle a blazing number of "ordinary" and AI inferenced transactions per second. Then there's the SAP HANA where entire workloads have data fully stored in the gargantuan RAM. Could I be neglecting to foresee certain trends in technology or industry? Perhaps. What I'm doing here is highlighting that the current niche is well served by POWER and Z. I don't see the current users switching entirely to Cloud - that ship sailed 5 years ago.
- jacquesm 3y ago
  Good points. Essentially Linux is the delivery mechanism for the power potential of the mainframes so that Linux is a convenience and not a sign that the tables are turning. It also helps to be able to run a mountain of software without having to develop it first for some proprietary OS. This is where the big difference between transaction processing software and massive parallel problems such as simulations lies. A bank doesn't need a supercomputer, it needs a reliable and scalable transaction processor. Off topic, and I'm sure you are not in a position to comment, but I find the 'Watson' marketing ploy a massive turn-off, it is so tone-deaf that I really don't understand why they keep pushing it so hard.
  - datameta 3y ago
    I'll be honest I haven't seeked out how Watson X is being projected externally, and don't know much about the details internally - it's being driven by a different division of the company. So even if I were in the position to comment, I don't believe I have anything meaningful to say there. You touched on a point I meant to add - reliability. Something that sets the Z mainframe apart from the Power server is the unyielding quest for absolutely no downtime (in exchange for compromise in raw compute/memory per drawer vs Power). I believe the quoted number is 3 seconds per year maximum. Because having huge throughput is well and good for business, but lost transactions could really strike at the bottom line with an unpredictable ripple effect. And in the case that a network of major hospitals is using the mainframe - no downtime could potentially save lives.
    - jacquesm 3y ago
      Exactly. I worked for a bank (using IBM hardware, as well as Sperry) in the good old days and back then the reliability factor is what drove each and every major purchase. Tandem made their whole business case around that. When the computers and the IT department surrounding them are a small rounding error on the total budget of your organization it makes sense to go for a solution like that and I guess there will always be customers that need that. Consider me convinced: IBM will still be around in 2035, or at least, the mainframe division will be. It also puts their RedHat acquisition in a different light, so thank you for that bit of insight.

But half the top supercomputers are a cluster of Linux x86 boxes, and the other half are Linux on some other architecture.

by comex 3y ago

But half the top supercomputers are a cluster of Linux x86 boxes, and the other half are Linux on some other architecture.

datameta 3y ago

I think it's not accurate to call their architecture a cluster of machines. Either way I'm not referring to Linux as an OS - it is really the only choice for enterprise scale computing. Regardless of what architectures supercomputers use, their use cases are different from that of server/mainframe high performance compute. 8 cores with on-chip AI acceleration running at 5GHz sharing virtual cache, using ~256GB-1TB of of RAM at ~76-304Gbps transfer speed is something a major bank would use. It won't cut it for SoTA atmospheric modeling, molecular modeling, or something like rapid computational chemistry search space exploration. What it excels at is the ability to handle a blazing number of "ordinary" and AI inferenced transactions per second. Then there's the SAP HANA where entire workloads have data fully stored in the gargantuan RAM. Could I be neglecting to foresee certain trends in technology or industry? Perhaps. What I'm doing here is highlighting that the current niche is well served by POWER and Z. I don't see the current users switching entirely to Cloud - that ship sailed 5 years ago.

jacquesm 3y ago
Good points. Essentially Linux is the delivery mechanism for the power potential of the mainframes so that Linux is a convenience and not a sign that the tables are turning. It also helps to be able to run a mountain of software without having to develop it first for some proprietary OS. This is where the big difference between transaction processing software and massive parallel problems such as simulations lies. A bank doesn't need a supercomputer, it needs a reliable and scalable transaction processor. Off topic, and I'm sure you are not in a position to comment, but I find the 'Watson' marketing ploy a massive turn-off, it is so tone-deaf that I really don't understand why they keep pushing it so hard.
- datameta 3y ago
  I'll be honest I haven't seeked out how Watson X is being projected externally, and don't know much about the details internally - it's being driven by a different division of the company. So even if I were in the position to comment, I don't believe I have anything meaningful to say there. You touched on a point I meant to add - reliability. Something that sets the Z mainframe apart from the Power server is the unyielding quest for absolutely no downtime (in exchange for compromise in raw compute/memory per drawer vs Power). I believe the quoted number is 3 seconds per year maximum. Because having huge throughput is well and good for business, but lost transactions could really strike at the bottom line with an unpredictable ripple effect. And in the case that a network of major hospitals is using the mainframe - no downtime could potentially save lives.
  - jacquesm 3y ago
    Exactly. I worked for a bank (using IBM hardware, as well as Sperry) in the good old days and back then the reliability factor is what drove each and every major purchase. Tandem made their whole business case around that. When the computers and the IT department surrounding them are a small rounding error on the total budget of your organization it makes sense to go for a solution like that and I guess there will always be customers that need that. Consider me convinced: IBM will still be around in 2035, or at least, the mainframe division will be. It also puts their RedHat acquisition in a different light, so thank you for that bit of insight.