Re: [Crash-utility] [PATCH 0/5] [RFC] Multi-thread support for search cmd

> > To be honest, I'm not sure if it's really worth introducing multi-thread only for a "search" > command, as the "search" command is not commonly used, and the performance issue > only occurs when the "search" command reads large memory. Let's see if Kazu has any > comments about it. > > But the [PATCH 5/5] looks good to me, and it should be a separate patch. > Thanks for your comments. The original motivation for patch5/5 is to make a thread-safe version of readmem(), which failed, but the patch5/5 itself is good to simplify the reading process. As for the

search cmd more often, and they also complain about other crash cmds such as kmem suffering from performance issues. They are still trying

gains and the cost of code maintenance doesn't have a good balance if only for one search cmd. I will see if the multi-thread code can work as a common library so other commands can benefit from it as well.

Anyway I want to see how maintainers think about the performance issues, whether multithread is an acceptable solution, and if there are any suggestions if I want to go further. Thanks, Tao Liu > Thanks. > Lianbo > >> >> steps: 1) readmem data into pagebuf, 2) search specific values within the >> pagebuf. A typical workflow of search is as follows: >> >> for addr from low to high: >> do >> readmem(addr, pagebuf) >> search_value(value, pagebuf) >> addr += pagesize >> done >> >> There are 2 points which we can accelerate: 1) readmem don't have to wait >> search_value, when search_value is working, readmem can read the next pagebuf >> at the same time. 2) higher addr don't have to wait lower addr, they can be >> processed at the same time if we carefully arrange the output order. >> >> For point 1, we introduce zones for pagebuf, e.g. search_value can work on >> zone 0 while readmem can prepare the data for zone 1. For point 2, we introduce >> multiple search_value in threads, e.g. readmem will prepare 100 pages as a >> batch, then we will have 4 threads of search_value, thread 0 handles page 1~25, >> thread 2 handles page 26~50 page, thread 3 handles page 51~75, thread 4 handles >> page 76~100. >> >> A typical workflow of multithread search implemented in this patchset is as >> follows, which removed thread synchronization: >> >> pagebuf[ZONE][BATCH] >> zone_index = buf_index = 0 >> create_thread(4, search_value) >> for addr from low to high: >> do >> if buf_index < BATCH >> readmem(addr, pagebuf[zone_index][buf_index++]) >> addr += pagesize >> else >> start_thread(pagebuf[zone_index], 0/4 * BATCH, 1/4 * BATCH) >> start_thread(pagebuf[zone_index], 1/4 * BATCH, 2/4 * BATCH) >> start_thread(pagebuf[zone_index], 2/4 * BATCH, 3/4 * BATCH) >> start_thread(pagebuf[zone_index], 3/4 * BATCH, 4/4 * BATCH) >> zone_index++ >> buf_index = 0 >> fi >> done >> >> readmem works in the main process and not multi-threaded, because readmem will >> not only read data from vmcore, decompress it, but walk through page tables if >> virtual address given. It is hard to reimplement it into thread safe version, >> search_value is easier to be made thread-safe. By carefully choose batch size >> and thread num, we can maximize the concurrency. >> >> The last part of the patchset, is replacing lseek/read to pread for kcore and >> diskdumped vmcore. >> >> Here is the performance test result chart. Please note the vmcore and >> kcore are tested seperately on 2 different machines. crash-orig is the >> crash compiled with clean upstream code, crash-pread is the code with only >> pread patch applied(patch 5), crash-multi is the code with only multithread >> patches applied(patch 1~4). >> >> ulong search: >> >> $ time echo "search abcd" | ./crash-orig vmcore vmlinux > /dev/null >> $ time echo "search abcd -f 4 -n 4" | ./crash-multi vmcore vmlinux > /dev/null >> >> 45G vmcore 64G kcore >> real user sys real user sys >> crash-orig 16m56.595s 15m57.188s 0m56.698s 1m37.982s 0m51.625s 0m46.266s >> crash-pread 16m46.366s 15m55.790s 0m48.894s 1m9.179s 0m36.646s 0m32.368s >> crash-multi 16m26.713s 19m8.722s 1m29.263s 1m27.661s 0m57.789s 0m54.604s >> >> string search: >> >> $ time echo "search -c abcddbca" | ./crash-orig vmcore vmlinux > /dev/null >> $ time echo "search -c abcddbca -f 4 -n 4" | ./crash-multi vmcore vmlinux > /dev/null >> >> 45G vmcore 64G kcore >> real user sys real user sys >> crash-orig 33m33.481s 32m38.321s 0m52.771s 8m32.034s 7m50.050s 0m41.478s >> crash-pread 33m25.623s 32m35.019s 0m47.394s 8m4.347s 7m35.352s 0m28.479s >> crash-multi 16m31.016s 38m27.456s 1m11.048s 5m11.725s 7m54.224s 0m44.186s >> >> Discussion: >> >> 1) Either multithread and pread patches can improve the performance a >> bit, so if both patches applied, the performance can be better. >> >> 2) Multi-thread search performs much better in search time consumptive >> tasks, such as string search. >> >> Tao Liu (5): >> Introduce multi-thread to search_virtual >> Introduce multi-thread to search_physical >> Introduce multi-thread to string search >> Introduce multi-thread options to search cmd >> Replace lseek/read into pread for kcore and vmcore reading. >> >> defs.h | 6 + >> diskdump.c | 11 +- >> help.c | 17 +- >> memory.c | 1176 +++++++++++++++++++++++++++++++++++++++++----------- >> netdump.c | 5 +- >> task.c | 14 + >> 6 files changed, 969 insertions(+), 260 deletions(-) >> >> -- >> 2.33.1