Not sure how many people know about the Dave Cutler videos that popped up recently. Here are four:
David Cutler — 2016 CHM Fellow
https://www.youtube.com/watch?v=J2GV_bCfnCw ( 16m )
Oral History of Dave Cutler Part 1 - the Computer History Museum (2016-02-25)
https://www.youtube.com/watch?v=29RkHH-psrY ( 01:44:33 )
Oral History of Dave Cutler Part 2 - the Computer History Museum (2016-02-25)
https://www.youtube.com/watch?v=SVgSLud50ss ( 01:21:18 )
The Mind Behind Windows: Dave Cutler - Dave's Garage (2023-10-21)
https://www.youtube.com/watch?v=xi1Lq79mLeE ( 03:10:22 )
Now being a true glutton for punishment, I watched them all. In one of the videos, Dave mentions that no Itanium machine was ever as fast as x86-64 available at the same time. Which got me thinking, our rx2600-i2 is now 9 years old, and I just came into possession of two used HPE DL380p_gen9 chassis (these things are really slick), Anyway, sometime soon, I'm going to acquire a trial copy of "OpenVMS x86-64" then see how these puppies behave. If all goes well, I am hoping that the bean-counters approve the purchase. That will mean that I've taken our bread-n-butter homegrown system from VAX, to Alpha, to Itanium, to x86-64. Stay tuned.
Dave Cutler speaks
Re: Dave Cutler speaks
I haven't watched the first video on your list yet but I have watched the other three, and they're extremely informative for someone interested in both DEC and Windows NT history. Robert Supnik's Computer History Museum oral history interviews fill in a lot of other fascinating DEC history from the years after Cutler had left DEC for Microsoft.
If not for Cutler's interviews, I wouldn't have known how influential Microsoft had been in terms of giving AMD their blessing to extend x86 to 64 bits, and apparently Microsoft's engineers had a few technical suggestions that AMD took up (easy long branches was something he mentioned, IIRC).
Right now I'm watching Mark Russinovich give a talk in 2004 about the Windows NT vs. Linux kernels:
https://youtu.be/HdV9QuvgS_w?si=tJ_m9lWrkUfoOhml
When it comes to Itanium vs. x86-64 for VMS specifically, everything's better on the x86 side except possibly the architectural details of the complexities of emulating the VAX-oriented requirements provided on Itanium and x86 by the SWIS layer. Camiel Vanderhoeven has some great talks about how that all works under the hood, which I assume are still accurate.
It sounds like the most critical CPU feature to have available is the Process-context identifier (PCID) address space extension, which has been around since 2010. Hopefully it works equally well inside any hypervisor you intend on using. Unlike Linux, which has added support for PCID for security (isolation from Meltdown attacks) and performance, VMS relies on being able to switch between 4 sets of page tables per process quickly, to emulate VAX's 4 rings of security per memory page.
I would set very conservative expectation with your bosses as far as the current x86 release's performance for anything I/O, filesystem, or network-intensive, because the LLVM-based compilers are so new and not all the optimization passes are enabled, the SWIS layer I mentioned is new, the OS executive and libraries haven't had time to be optimized for the CPU yet, and there are some fairly serious known issues with the current "X" version compilers. One example: I got some of the basic COBOL 85 tests to hang when I tried running them (using the GnuCOBOL scripts on another system to extract the .CBL files from the archive), and some of the output was going to the wrong place. The release notes for both COBOL and Fortran say that some overflow cases aren't handled yet.
Since you already have hardware and can presumably get access to licenses for evaluation and testing, it seems like it's well worth your time to try it out. I just wouldn't get my hopes too high as far as immediately blowing away your current Itanium setup, depending on your workload. The Java and C++ compilers are probably most improved over Itanium. Anything that goes through the GEM->LLVM pass is still very new.
If not for Cutler's interviews, I wouldn't have known how influential Microsoft had been in terms of giving AMD their blessing to extend x86 to 64 bits, and apparently Microsoft's engineers had a few technical suggestions that AMD took up (easy long branches was something he mentioned, IIRC).
Right now I'm watching Mark Russinovich give a talk in 2004 about the Windows NT vs. Linux kernels:
https://youtu.be/HdV9QuvgS_w?si=tJ_m9lWrkUfoOhml
When it comes to Itanium vs. x86-64 for VMS specifically, everything's better on the x86 side except possibly the architectural details of the complexities of emulating the VAX-oriented requirements provided on Itanium and x86 by the SWIS layer. Camiel Vanderhoeven has some great talks about how that all works under the hood, which I assume are still accurate.
It sounds like the most critical CPU feature to have available is the Process-context identifier (PCID) address space extension, which has been around since 2010. Hopefully it works equally well inside any hypervisor you intend on using. Unlike Linux, which has added support for PCID for security (isolation from Meltdown attacks) and performance, VMS relies on being able to switch between 4 sets of page tables per process quickly, to emulate VAX's 4 rings of security per memory page.
I would set very conservative expectation with your bosses as far as the current x86 release's performance for anything I/O, filesystem, or network-intensive, because the LLVM-based compilers are so new and not all the optimization passes are enabled, the SWIS layer I mentioned is new, the OS executive and libraries haven't had time to be optimized for the CPU yet, and there are some fairly serious known issues with the current "X" version compilers. One example: I got some of the basic COBOL 85 tests to hang when I tried running them (using the GnuCOBOL scripts on another system to extract the .CBL files from the archive), and some of the output was going to the wrong place. The release notes for both COBOL and Fortran say that some overflow cases aren't handled yet.
Since you already have hardware and can presumably get access to licenses for evaluation and testing, it seems like it's well worth your time to try it out. I just wouldn't get my hopes too high as far as immediately blowing away your current Itanium setup, depending on your workload. The Java and C++ compilers are probably most improved over Itanium. Anything that goes through the GEM->LLVM pass is still very new.
-
neilrieck
Topic author - Contributor
- Posts: 21
- Joined: Tue Jan 10, 2023 10:41 am
- Reputation: 0
- Location: Waterloo, Ontario, Canada
- Status: Offline
- Contact:
Re: Dave Cutler speaks
Me as well. In one video, Cutler mentions seeing a pre-release of AMD's 64-bit extensions to the 32-bit x86 before AMD went public. He mentions recommending changes like "a change to register-based indirect addressing which would simplify implementing PIC (position independent code)" which AMD implemented, then Intel was required to copy. This whole thing reminded me of Cutler's role at DEC West (Seattle) where he was managing the Prism and Mica projects.
https://en.wikipedia.org/wiki/DEC_PRISM
https://en.wikipedia.org/wiki/DEC_MICA
Anyway, I wonder if this unsung hero is one of the primary reasons why x86-64 is found almost every. I wonder what he thinks about Apple's latest love affair with ARM processors.
https://en.wikipedia.org/wiki/DEC_PRISM
https://en.wikipedia.org/wiki/DEC_MICA
Anyway, I wonder if this unsung hero is one of the primary reasons why x86-64 is found almost every. I wonder what he thinks about Apple's latest love affair with ARM processors.