[PATCH v4 2/2] Use the POWER8 Micro Partition Prefetch Engine in KVM HV on POWER8

[Paul Mackerras]

On Fri, Jul 18, 2014 at 02:18:43PM +1000, Stewart Smith wrote:
> The POWER8 processor has a Micro Partition Prefetch Engine, which is
> a fancy way of saying "has way to store and load contents of L2 or
> L2+MRU way of L3 cache". We initiate the storing of the log (list of
> addresses) using the logmpp instruction and start restore by writing
> to a SPR.
> 
> The logmpp instruction takes parameters in a single 64bit register:
> - starting address of the table to store log of L2/L2+L3 cache contents
>   - 32kb for L2
>   - 128kb for L2+L3
>   - Aligned relative to maximum size of the table (32kb or 128kb)
> - Log control (no-op, L2 only, L2 and L3, abort logout)
> 
> We should abort any ongoing logging before initiating one.
> 
> To initiate restore, we write to the MPPR SPR. The format of what to write
> to the SPR is similar to the logmpp instruction parameter:
> - starting address of the table to read from (same alignment requirements)
> - table size (no data, until end of table)
> - prefetch rate (from fastest possible to slower. about every 8, 16, 24 or
>   32 cycles)
> 
> The idea behind loading and storing the contents of L2/L3 cache is to
> reduce memory latency in a system that is frequently swapping vcores on
> a physical CPU.
> 
> The best case scenario for doing this is when some vcores are doing very
> cache heavy workloads. The worst case is when they have about 0 cache hits,
> so we just generate needless memory operations.
> 
> This implementation just does L2 store/load. In my benchmarks this proves
> to be useful.
> 
> Benchmark 1:
>  - 16 core POWER8
>  - 3x Ubuntu 14.04LTS guests (LE) with 8 VCPUs each
>  - No split core/SMT
>  - two guests running sysbench memory test.
>    sysbench --test=memory --num-threads=8 run
>  - one guest running apache bench (of default HTML page)
>    ab -n 490000 -c 400 http://localhost/
> 
> This benchmark aims to measure performance of real world application (apache)
> where other guests are cache hot with their own workloads. The sysbench memory
> benchmark does pointer sized writes to a (small) memory buffer in a loop.
> 
> In this benchmark with this patch I can see an improvement both in requests
> per second (~5%) and in mean and median response times (again, about 5%).
> The spread of minimum and maximum response times were largely unchanged.
> 
> benchmark 2:
>  - Same VM config as benchmark 1
>  - all three guests running sysbench memory benchmark
> 
> This benchmark aims to see if there is a positive or negative affect to this
> cache heavy benchmark. Although due to the nature of the benchmark (stores) we
> may not see a difference in performance, but rather hopefully an improvement
> in consistency of performance (when vcore switched in, don't have to wait
> many times for cachelines to be pulled in)
> 
> The results of this benchmark are improvements in consistency of performance
> rather than performance itself. With this patch, the few outliers in duration
> go away and we get more consistent performance in each guest.
> 
> benchmark 3:
>  - same 3 guests and CPU configuration as benchmark 1 and 2.
>  - two idle guests
>  - 1 guest running STREAM benchmark
> 
> This scenario also saw performance improvement with this patch. On Copy and
> Scale workloads from STREAM, I got 5-6% improvement with this patch. For
> Add and triad, it was around 10% (or more).
> 
> benchmark 4:
>  - same 3 guests as previous benchmarks
>  - two guests running sysbench --memory, distinctly different cache heavy
>    workload
>  - one guest running STREAM benchmark.
> 
> Similar improvements to benchmark 3.
> 
> benchmark 5:
>  - 1 guest, 8 VCPUs, Ubuntu 14.04
>  - Host configured with split core (SMT8, subcores-per-core=4)
>  - STREAM benchmark
> 
> In this benchmark, we see a 10-20% performance improvement across the board
> of STREAM benchmark results with this patch.
> 
> Based on preliminary investigation and microbenchmarks
> by Prerna Saxena <prerna at linux.vnet.ibm.com>
> 
> Signed-off-by: Stewart Smith <stewart at linux.vnet.ibm.com>

Acked-by: Paul Mackerras <paulus at samba.org>