I use perf top and try to see where in the kernel code all this time was being spend. Following is what I was seeing: PerfTop: 2036 irqs/sec kernel:100.0% exact: 0.0% [1000Hz cycles], (target_pid: 5336) ---------------------------------------------------------------------------------------------------------- samples pcnt function DSO _______ _____ _________________________ _________________ 11265.00 91.0% _spin_lock_irq [kernel.kallsyms] 355.00 2.9% _spin_lock_irqsave [kernel.kallsyms] 249.00 2.0% compaction_alloc [kernel.kallsyms] 235.00 1.9% compact_zone [kernel.kallsyms] 151.00 1.2% get_pageblock_flags_group [kernel.kallsyms] 32.00 0.3% _cond_resched [kernel.kallsyms] 27.00 0.2% copy_page_c [kernel.kallsyms] 8.00 0.1% _spin_lock [kernel.kallsyms] 6.00 0.0% mem_cgroup_del_lru_list [kernel.kallsyms] 5.00 0.0% __wake_up_bit [kernel.kallsyms] Then, I use “perf record -g -p 5336” to capture the percent of kernel call and find “compact_zone” is very busy. [root@bj-nx-cip-w87 ~]# perf report --stdio # Events: 47K cycles # # Overhead Command Shared Object Symbol # ........ ......... ..................... ...................................... # 91.51% glusterfs [kernel.kallsyms] [k] _spin_lock_irq | --- _spin_lock_irq | |--99.64%-- compact_zone | compact_zone_order | try_to_compact_pages | __alloc_pages_nodemask | alloc_pages_vma | do_huge_pmd_anonymous_page | handle_mm_fault | __do_page_fault | do_page_fault | page_fault | | | |--94.45%-- xdr_callmsg_internal | | 0x3829b98860 | | | --5.55%-- __memcpy_sse2 --0.36%-- [...] Last, I google “compact_zone” and find the article “Linux 6 Transparent Huge Pages and Hadoop Workloads”. The sample issues is occurred in hadoop. THP can be disabled by running the following command: echo never > /sys/kernel/mm/redhat_transparent_hugepage/enabled CPU usage and work load is normal. From: gluster-devel-bounces+gluster=163.com@xxxxxxxxxx [mailto:gluster-devel-bounces+gluster=163.com@xxxxxxxxxx] On Behalf Of Song We have met same performance issues when open a file. Sometime, it cost more than 10 seconds that a file open. We add some debug info to locate this problem and test again and again, find it's will cost a few seconds when execute 'xdr_callmsg' function in 'rpc_request_to_xdr'. A typical example of log is as below. From [2013-06-06 13:34:59.471004] to [2013-06-06 13:35:04.890363] , 'xdr_callmsg' function cost more than 5 seconds. [2013-06-06 13:34:59.470991] I [rpc-clnt.c:1175:rpc_clnt_record_build_record] 0-gfs1-client-51: (thread_id is 140257410492160 )add for open_slow rpc_fill_request_end [2013-06-06 13:34:59.471004] I [xdr-rpcclnt.c:87:rpc_request_to_xdr] 0-rpc: (thread_id is 140257410492160 len = 131072 )add for open_slow xdrmem_create_end [2013-06-06 13:34:59.570044] I [client.c:124:client_submit_request] 0-gfs1-client-86: (thread_id is 140257819739904 )add for open_slow rpc_clnt_submit [2013-06-06 13:34:59.570091] I [rpc-clnt.c:1363:rpc_clnt_submit] 0-gfs1-client-86: (thread_id is 140257819739904 )add for open_slow callid end ...... [2013-06-06 13:34:59.579865] I [client3_1-fops.c:2235:client3_1_lookup_cbk] 0-gfs1-client-5: (thread_id is 140257819739904)add for open_slow lookup_cbk path=/xmail_dedup/gfs1_000/1FA/1B1 [2013-06-06 13:34:59.579917] I [client3_1-fops.c:2235:client3_1_lookup_cbk] 0-gfs1-client-6: (thread_id is 140257819739904)add for open_slow lookup_cbk path=/xmail_dedup/gfs1_000/1FA/1B1 [2013-06-06 13:35:04.890363] I [xdr-rpcclnt.c:92:rpc_request_to_xdr] 0-rpc: (thread_id is 140257410492160 )add for open_slow xdr_callmsg_end [2013-06-06 13:35:04.890366] I [client.c:110:client_submit_request] 0-gfs1-client-44: (thread_id is 140257785079552 )add for open_slow create the xdr payload Native client and use 5 glusterfs in one server. When performance issues appear, the cpu usage is as below: top - 13:45:37 up 57 days, 14:04, 4 users, load average: 6.98, 5.38, 4.67 Tasks: 712 total, 8 running, 704 sleeping, 0 stopped, 0 zombie Cpu(s): 3.2%us, 63.5%sy, 0.0%ni, 31.5%id, 1.4%wa, 0.0%hi, 0.4%si, 0.0%st Mem: 65956748k total, 55218008k used, 10738740k free, 3362972k buffers Swap: 8388600k total, 41448k used, 8347152k free, 37370840k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 13905 root 20 0 554m 363m 2172 R 244.8 0.6 9:51.01 glusterfs 13650 root 20 0 766m 610m 2056 R 184.8 0.9 18:24.37 glusterfs 13898 root 20 0 545m 356m 2176 R 179.2 0.6 12:04.87 glusterfs 13919 root 20 0 547m 360m 2172 R 111.6 0.6 9:16.89 glusterfs 22460 root 20 0 486m 296m 2200 S 100.4 0.5 194:59.10 glusterfs 13878 root 20 0 545m 361m 2176 R 99.7 0.6 8:35.88 glusterfs -----Original Message----- On Thu, 06 Jun 2013 10:39:21 -0300 Pablo <paa.listas@xxxxxxxxx> wrote: > I have never try this (In fact I'm just learning a bit more how to > administer a Gluster server.), buy you may find it useful. > > http://download.gluster.org/pub/gluster/glusterfs/doc/HA%20and%20Load% > 20Balancing%20for%20NFS%20and%20SMB.html > > Pablo. The thing with this way of failover is though, that you will likely corrupt a currently written file. If your NFS-server (gluster) node dies while you write your file will be corrupt. If you use native glusterfs mounts it will not (should not). This is why I consider the NFS server feature nothing more than a bad hack. It does not deliver the safety that glusterfs promises, even if you solve the failover problem somehow. -- Regards, Stephan _______________________________________________ Gluster-users mailing list http://supercolony.gluster.org/mailman/listinfo/gluster-users |
