5 ms·

I'm not sure how much water the multi tenant HA argument holds, these institutions are already doing that on AWS / Azure / k8s for the more modern apps. Rewrit

I'm not sure how much water the multi tenant HA argument holds, these institutions are already doing that on AWS / Azure / k8s for the more modern apps. Rewriting all the old crusty COBOL code might be a different story.

dralley 3y ago
Getting the same performance out of AWS as a local mainframe is neither trivial nor cheap for a lot of the workloads these systems are running.
- jeffbee 3y ago
  Getting stupefying amounts of I/O on AWS is as simple as turning a knob and getting out your credit card. The kind of throughput quoted in the article—48GB/s—is achievable on AWS for $250k/year, list price.
  - ElectricalUnion 3y ago
    I thought such throughput would cost around 30 274 560 USD/year (60 * 60 * 24 * 365 * 48 * 0.02) on data transit fees alone.
    - jeffbee 3y ago
      Obviously it would be very foolish to use an AWS storage backend for an application that was not also in AWS. That said, it is not at all clear to me that the TCO of a maxed-out IBM DS8900F is less than $30 million/year.
  - dralley 3y ago
    But let's say you want that data compressed and/or encrypted, you're going to need an awful lot of compute to exploit that 48GB/s if it's possible at all. Whereas on mainframe the service processors can deal with it transparently to the CPU using specialized hardware. And let's say it's financial transactions - fixed-point decimal arithmetic. On mainframe, that can be hardware accelerated too.
    - jeffbee 3y ago
      Yes, and it's interesting how Intel is moving in that direction with their Xeon line, integrating the QAT peripheral for compression and encryption offloading. Unfortunately these things are unicorns and that aren't generally available in the cloud, either.
      - ghaff 3y ago
        Well, the general approach over the past 30 years or so has just been to mostly go with generic x86 hardware which would get predictably faster year-over-year and software could be written pretty generically for those platforms. With performance growth slowing, one of the answers has been to have more heterogenous and special function hardware and the software people will just have to deal with it.
  - shrubble 3y ago
    That is one channel of throughput, and a mainframe can have many channels, though.
- WJW 3y ago
  Getting (and maintaining) a local mainframe in the first place is also not cheap though. I won't hazard a guess to which one might be easier to keep up and running, but at least in my bubble it would be 100x easier to hire more k8s admins than to hire qualified mainframe admins.

dralley 3y ago

Getting the same performance out of AWS as a local mainframe is neither trivial nor cheap for a lot of the workloads these systems are running.

jeffbee 3y ago
Getting stupefying amounts of I/O on AWS is as simple as turning a knob and getting out your credit card. The kind of throughput quoted in the article—48GB/s—is achievable on AWS for $250k/year, list price.
- ElectricalUnion 3y ago
  I thought such throughput would cost around 30 274 560 USD/year (60 * 60 * 24 * 365 * 48 * 0.02) on data transit fees alone.
  - jeffbee 3y ago
    Obviously it would be very foolish to use an AWS storage backend for an application that was not also in AWS. That said, it is not at all clear to me that the TCO of a maxed-out IBM DS8900F is less than $30 million/year.
- dralley 3y ago
  But let's say you want that data compressed and/or encrypted, you're going to need an awful lot of compute to exploit that 48GB/s if it's possible at all. Whereas on mainframe the service processors can deal with it transparently to the CPU using specialized hardware. And let's say it's financial transactions - fixed-point decimal arithmetic. On mainframe, that can be hardware accelerated too.
  - jeffbee 3y ago
    Yes, and it's interesting how Intel is moving in that direction with their Xeon line, integrating the QAT peripheral for compression and encryption offloading. Unfortunately these things are unicorns and that aren't generally available in the cloud, either.
    - ghaff 3y ago
      Well, the general approach over the past 30 years or so has just been to mostly go with generic x86 hardware which would get predictably faster year-over-year and software could be written pretty generically for those platforms. With performance growth slowing, one of the answers has been to have more heterogenous and special function hardware and the software people will just have to deal with it.
- shrubble 3y ago
  That is one channel of throughput, and a mainframe can have many channels, though.
WJW 3y ago
Getting (and maintaining) a local mainframe in the first place is also not cheap though. I won't hazard a guess to which one might be easier to keep up and running, but at least in my bubble it would be 100x easier to hire more k8s admins than to hire qualified mainframe admins.