On Fri, 4 Sep 2020 09:22:09 -0700, Ed Jaffe <edja...@phoenixsoftware.com> wrote:
>IMHO, if you need additional zIIP capacity for a production workload, it >probably makes more sense to configure another zIIP core online than it >does to enable MT=2. Agreed. SMT is a good thing to keep in your pocket for the emergency of "we need a little more zIIP capacity, but we can't purchase another zIIP right now". Don't use it as a substitute for real zIIP capacity. If you need zIIP capacity buy more zIIP capacity. It's relatively inexpensive. 1) The amount of extra throughput you'll get is variable depending on your workload mix. And the timing of the mix of workload. You can't use test to accurately predict production, you just have to try it. 2) MT=2 is effectively more/slower engines vs. fewer/faster engines. This is a trade-off that's ok to good for most workloads, but is worth bearing in mind. 3) The measurements are complicated estimates based on instruction counts seen with 1 and 2 threads active that are not well documented. Most importantly, the reported zIIP time becomes MT1ET: the estimated time the work would have consumed on the zIIP had the zIIP been running in MT=1 instead of MT=2. But remember it's an estimate and based on numbers that change based on the instruction mix. Doing accurate capacity planning with estimates that vary this way is difficult. Which may not be a problem if your capacity planning for zIIPs is "when we start seeing cross-over to the GCPs we buy more zIIPs" (which is not necessarily a bad policy). SMT is a useful tool in certain situations, but IMHO, I wouldn't consider enabling MT=2 as the default. If you're having a problem that more zIIP capacity might help, then more real zIIP capacity is the best answer. In cases where that's impractical, MT=2 might be useful. Scott Chapman ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
