Re: [PATCH v3] pseries/hotplug-memory: hot-add: skip redundant LMB lookup
From: Scott Cheloha <hidden>
Date: 2020-09-16 14:43:35
On Wed, Sep 16, 2020 at 09:39:53AM +0200, David Hildenbrand wrote:
On 15.09.20 21:46, Scott Cheloha wrote:quoted
During memory hot-add, dlpar_add_lmb() calls memory_add_physaddr_to_nid() to determine which node id (nid) to use when later calling __add_memory(). This is wasteful. On pseries, memory_add_physaddr_to_nid() finds an appropriate nid for a given address by looking up the LMB containing the address and then passing that LMB to of_drconf_to_nid_single() to get the nid. In dlpar_add_lmb() we get this address from the LMB itself. In short, we have a pointer to an LMB and then we are searching for that LMB *again* in order to find its nid. If we call of_drconf_to_nid_single() directly from dlpar_add_lmb() we can skip the redundant lookup. The only error handling we need to duplicate from memory_add_physaddr_to_nid() is the fallback to the default nid when drconf_to_nid_single() returns -1 (NUMA_NO_NODE) or an invalid nid. Skipping the extra lookup makes hot-add operations faster, especially on machines with many LMBs. Consider an LPAR with 126976 LMBs. In one test, hot-adding 126000 LMBs on an upatched kernel took ~3.5 hours while a patched kernel completed the same operation in ~2 hours: Unpatched (12450 seconds): Sep 9 04:06:31 ltc-brazos1 drmgr[810169]: drmgr: -c mem -a -q 126000 Sep 9 04:06:31 ltc-brazos1 kernel: pseries-hotplug-mem: Attempting to hot-add 126000 LMB(s) [...] Sep 9 07:34:01 ltc-brazos1 kernel: pseries-hotplug-mem: Memory at 20000000 (drc index 80000002) was hot-added Patched (7065 seconds): Sep 8 21:49:57 ltc-brazos1 drmgr[877703]: drmgr: -c mem -a -q 126000 Sep 8 21:49:57 ltc-brazos1 kernel: pseries-hotplug-mem: Attempting to hot-add 126000 LMB(s) [...] Sep 8 23:27:42 ltc-brazos1 kernel: pseries-hotplug-mem: Memory at 20000000 (drc index 80000002) was hot-added It should be noted that the speedup grows more substantial when hot-adding LMBs at the end of the drconf range. This is because we are skipping a linear LMB search. To see the distinction, consider smaller hot-add test on the same LPAR. A perf-stat run with 10 iterations showed that hot-adding 4096 LMBs completed less than 1 second faster on a patched kernel: Unpatched: Performance counter stats for 'drmgr -c mem -a -q 4096' (10 runs): 104,753.42 msec task-clock # 0.992 CPUs utilized ( +- 0.55% ) 4,708 context-switches # 0.045 K/sec ( +- 0.69% ) 2,444 cpu-migrations # 0.023 K/sec ( +- 1.25% ) 394 page-faults # 0.004 K/sec ( +- 0.22% ) 445,902,503,057 cycles # 4.257 GHz ( +- 0.55% ) (66.67%) 8,558,376,740 stalled-cycles-frontend # 1.92% frontend cycles idle ( +- 0.88% ) (49.99%) 300,346,181,651 stalled-cycles-backend # 67.36% backend cycles idle ( +- 0.76% ) (50.01%) 258,091,488,691 instructions # 0.58 insn per cycle # 1.16 stalled cycles per insn ( +- 0.22% ) (66.67%) 70,568,169,256 branches # 673.660 M/sec ( +- 0.17% ) (50.01%) 3,100,725,426 branch-misses # 4.39% of all branches ( +- 0.20% ) (49.99%) 105.583 +- 0.589 seconds time elapsed ( +- 0.56% ) Patched: Performance counter stats for 'drmgr -c mem -a -q 4096' (10 runs): 104,055.69 msec task-clock # 0.993 CPUs utilized ( +- 0.32% ) 4,606 context-switches # 0.044 K/sec ( +- 0.20% ) 2,463 cpu-migrations # 0.024 K/sec ( +- 0.93% ) 394 page-faults # 0.004 K/sec ( +- 0.25% ) 442,951,129,921 cycles # 4.257 GHz ( +- 0.32% ) (66.66%) 8,710,413,329 stalled-cycles-frontend # 1.97% frontend cycles idle ( +- 0.47% ) (50.06%) 299,656,905,836 stalled-cycles-backend # 67.65% backend cycles idle ( +- 0.39% ) (50.02%) 252,731,168,193 instructions # 0.57 insn per cycle # 1.19 stalled cycles per insn ( +- 0.20% ) (66.66%) 68,902,851,121 branches # 662.173 M/sec ( +- 0.13% ) (49.94%) 3,100,242,882 branch-misses # 4.50% of all branches ( +- 0.15% ) (49.98%) 104.829 +- 0.325 seconds time elapsed ( +- 0.31% ) This is consistent. An add-by-count hot-add operation adds LMBs greedily, so LMBs near the start of the drconf range are considered first. On an otherwise idle LPAR with so many LMBs we would expect to find the LMBs we need near the start of the drconf range, hence the smaller speedup. Signed-off-by: Scott Cheloha <redacted>Hi Scott, IIRC, ppc DLPAR does a single add_memory() [...]
Yes.
[...] for each LMB (16 MB).
The block size is set by the hypervisor. The default is 256MB. In this test I had a block size of 256MB. On multi-terabyte machines I would effectively always expect a block size of 256MB. 16MB blocks are supported, but it is not the default setting so it is increasingly rare.
With tons of LMBs, this will also make /proc/iomem explode in size (using a list-based tree), making traversal significantly slower e.g., on insertions and system ram walks. I was wondering if you would get another performance boost under ppc when using MEMHP_MERGE_RESOURCE [1]. AFAIKs, the resource boundaries are not of interest. No guarantees, might be worth a try.
I'll give it a shot.
Did you investigate what else makes memory hotplug that slow? (126000 LMBs correspond to roughly 2TB, that shouldn't take 2 hours ...)
It was about ~31TB in 256MB blocks. It's a worst-case test (add all the memory), but I'm pretty happy with a 1.5 hour improvement :)
Memory block devices might still be a slowdown (although we have an xarray in place now that takes care of most pain).
Memory block devices are no longer a hotspot. Some of the slowdown is in the printk overhead. We print a log for every LMB. It is very silly. I intend to move those to a debug priority, which should trivially speed things up. Otherwise I need to do more profiling.