Conversation
Longhorn
never_released@mastodon.social
Is it me or networking on IBM z is quite slow by modern standards
or seeing 8 800GbE NICs on one system spoiled me
1
0
0
Palmer Dabbelt
palmer
@never_released I'm not a mainframe guy, but IIUC a lot of what would go over the network on PC-derived servers goes over the interconnect on mainframes. So the network side of things there is really only for the traffic that needs to cross the boundary between the mainframe world and the regular world.
1
0
1
@palmer nowadays there are 192c server CPUs on a single socket (AMD Turin dense, AWS Graviton5 among others)
and a 4 drawer (full) config of a z17 tops out at a mere 200 cores
so looks like their historical advantages are heavily eroding fast/have already faded away
1
0
0
Palmer Dabbelt
palmer
@never_released still not really a mainframe guy, but IIUC they're not really about CPU performance and they haven't really been all that compelling at that in a long time.
I'm always a bit lost when trying to read through the IBM manuals, but looks like they still have a bigger single shared memory domain that Intel or AMD -- 64TiB on the IBM stuff, 32TiB on Intel and 12TiB on AMD. They're also got some crazy-sounding cache sizes (36MiB private L2, for example), and I think that's part of how the handle their huge systems (though the other systems seem to be catching up there, too).
I've also really never understood their I/O architecture, but IIUC they push a lot of stuff that PC-derived systems would do on the CPU into the I/O subsystem's co-processors and that's part of the mainframe magic -- though with modern I/O stuff I'm not sure how true that is any more, the high-performance stuff is pretty hands-off from the CPU side of things and has a decent amount of compute living on the other end of PCIe.
I'm always a bit lost when trying to read through the IBM manuals, but looks like they still have a bigger single shared memory domain that Intel or AMD -- 64TiB on the IBM stuff, 32TiB on Intel and 12TiB on AMD. They're also got some crazy-sounding cache sizes (36MiB private L2, for example), and I think that's part of how the handle their huge systems (though the other systems seem to be catching up there, too).
I've also really never understood their I/O architecture, but IIUC they push a lot of stuff that PC-derived systems would do on the CPU into the I/O subsystem's co-processors and that's part of the mainframe magic -- though with modern I/O stuff I'm not sure how true that is any more, the high-performance stuff is pretty hands-off from the CPU side of things and has a decent amount of compute living on the other end of PCIe.
0
0
1