Date: Mon, 2 Dec 2019 14:15:26 -0800 From: Mark Millard <marklmi@yahoo.com> To: freebsd-arm@freebsd.org Subject: Comparing the OverDrive 1000 (A57) vs. MACCHIATObin Double Shot (A72) for buildworld and via a CPU/cache/RAM tradeoff-exploring benchmark Message-ID: <92E7B63A-E790-4815-9D91-2161A4F66B71@yahoo.com> References: <92E7B63A-E790-4815-9D91-2161A4F66B71.ref@yahoo.com>
next in thread | previous in thread | raw e-mail | index | archive | help
It looks like the OverDrive 1000 vs. MACCHIATObin Double Shot comparison ends up being an example of memory access making the difference for the specific workload: -j4 buildworld for head -r355027 (building itself from scratch). buildworld times (not needing a llvm bootstrap build): OverDrive 1000: 13895 sec (about 3.86 hrs) MACCHIATObin Double Shot: 16561 sec (about 4.60 hrs) So a little under 45 min difference when the mean and geometric mean are both a little over 4.2 hrs. SSD ufs file systems: One with Samsung 860 Pro, the other with Samsung 850 Pro. I do not expect that I/O made much of a difference, but I did nothing to measure such for the buildworld activity. OverDrive RAM: 8GiByte, half in each of the 2 slots MACCHIATObin RAM: 16GiByte, all in its 1 slot. MACCHIATObin: jumpers set for the fastest CPU/RAM speed for the Double Shot. A comparison graph from exploring single threaded and multi-threaded CPU/cache and RAM limited performance (a variation on the old HINT serial and pthread benchmarks) is shown at: = https://github.com/markmi/acpphint/blob/master/acpphint_example_data/acpph= int-OverDrive_1000_MacchDblShot-threads_4-LP64-g%2B%2B_9_8.3_O3-libc%2B%2B= _libstdc%2B%2B-DSIZE_large_fast_types-RAM.gp There are curves for various involved types: double (d), unsigned long long (ull), unsigned long (ul), unsigned int (ui). The match for ull and ul for the context provides some evidence of the variability observed. (The OverDrive and MACCHIATObin were not benchmarked for the graph at the same version of head: -r352341 based vs. -r355027 based.) (I did not set things such that the benchmark run would explore paging getting involved. Thus there is basically no I/O considered in the comparison graph.) The MACCHIATObin clearly wins single threaded and its memory subsystem was well matched to the single threaded use when the same-invovled-types are compared. (Single threaded are the blueish curves, MACCHIATObin having the lighter colors.) For multi-threaded in the range where RAM access limits things, the two systems are a close match. (Greenish colors, right side of plot, upper curves.) The range were the OverDrive 1000 is clearly faster is part of the middle of the multi-threaded curves. (This might be tied to whatever is done with the dual RAM slot structure or to the amount of caching, or some such, I do not know the details.) I would expect "-j1 buildworld" would take less time on the MACCHIATObin than on the OverDrive, but I'm not planing on measuring that. A more historical comparison, old PowerMac11,2 (2 sockets, 2 cores each) vs. the MACCHIATObin, both having 16 GiBytes of RAM: For analogous benchmark graphs (matching types), the MACCHIATObin single threaded is faster than the old PowerMac11,2 single threaded and also is usually faster than that 11,2's multi-threaded benchmark data as well. Multi-threaded, the MACCHIATObin is faster for the exploration by the benchmark. = https://github.com/markmi/acpphint/blob/master/acpphint_example_data/acpph= int-MacchDblShot_PowerMac11%2C2-threads_4-LP64-g%2B%2B_9_O3-libc%2B%2B-DSI= ZE_large_fast_types-RAM.gp I expect that this is interesting for the likely difference in power usage during the benchmarking. (Not that I've measured the power usage.) (The FreeBSD head vintages are not the same in the graph: -r355027 based vs. -r352341 based.) =3D=3D=3D Mark Millard marklmi at yahoo.com ( dsl-only.net went away in early 2018-Mar)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?92E7B63A-E790-4815-9D91-2161A4F66B71>