samuelnair | 166bf0f | 2017-09-05 00:57:49 +0800 | [diff] [blame] | 1 | Demonstrations of nfsslower, the Linux eBPF/bcc version. |
| 2 | |
| 3 | nfsslower show NFS reads, writes, opens and getattrs, slower than a |
| 4 | threshold. For example: |
| 5 | |
| 6 | ./nfsslower.py |
| 7 | Tracing NFS operations that are slower than 10 ms |
| 8 | TIME COMM PID T BYTES OFF_KB LAT(ms) FILENAME |
| 9 | 11:25:16 dd 21295 W 1048576 15360 14.84 1.test |
| 10 | 11:25:16 dd 21295 W 1048576 16384 12.73 1.test |
| 11 | 11:25:16 dd 21295 W 1048576 17408 24.27 1.test |
| 12 | 11:25:16 dd 21295 W 1048576 18432 22.93 1.test |
| 13 | 11:25:16 dd 21295 W 1048576 19456 14.65 1.test |
| 14 | 11:25:16 dd 21295 W 1048576 20480 12.58 1.test |
| 15 | 11:25:16 dd 21297 W 1048576 6144 10.50 1.test.w |
| 16 | 11:25:16 dd 21297 W 1048576 7168 16.65 1.test.w |
| 17 | 11:25:16 dd 21297 W 1048576 8192 13.01 1.test.w |
| 18 | 11:25:16 dd 21297 W 1048576 9216 14.06 1.test.w |
| 19 | |
| 20 | This shows NFS writes from dd each 1MB in size to 2 different files. The |
| 21 | writes all had latency higher than 10ms. |
| 22 | |
| 23 | This "latency" is measured from when the operation was issued from the VFS |
| 24 | interface to the file system, to when it completed. This spans everything: |
| 25 | RPC latency, network latency, file system CPU cycles, file system locks, run |
| 26 | queue latency, etc. This is a better measure of the latency suffered by |
| 27 | applications reading from a NFS share and can better expose problems |
| 28 | experienced by NFS clients. |
| 29 | |
| 30 | Note that this only traces the common NFS operations (read,write,open and |
| 31 | getattr). I chose to include getattr as a significant percentage of NFS |
| 32 | traffic end up being getattr calls and are a good indicator of problems |
| 33 | with an NFS server. |
| 34 | |
| 35 | The threshold can be provided as an argument. E.g. I/O slower than 1 ms: |
| 36 | |
| 37 | ./nfsslower.py 1 |
| 38 | Tracing NFS operations that are slower than 1 ms |
| 39 | TIME COMM PID T BYTES OFF_KB LAT(ms) FILENAME |
| 40 | 11:40:16 cp 21583 R 131072 0 4.35 1.test |
| 41 | 11:40:16 cp 21583 R 131072 256 1.87 1.test |
| 42 | 11:40:16 cp 21583 R 131072 384 2.99 1.test |
| 43 | 11:40:16 cp 21583 R 131072 512 4.19 1.test |
| 44 | 11:40:16 cp 21583 R 131072 640 4.25 1.test |
| 45 | 11:40:16 cp 21583 R 131072 768 4.65 1.test |
| 46 | 11:40:16 cp 21583 R 131072 1280 1.08 1.test |
| 47 | 11:40:16 cp 21583 R 131072 1408 3.29 1.test |
| 48 | 11:40:16 cp 21583 R 131072 1792 3.12 1.test |
| 49 | 11:40:16 cp 21583 R 131072 3712 3.55 1.test |
| 50 | 11:40:16 cp 21583 R 131072 3840 1.12 1.test |
| 51 | 11:40:16 cp 21583 R 131072 4096 3.23 1.test |
| 52 | 11:40:16 cp 21583 R 131072 4224 2.73 1.test |
| 53 | 11:40:16 cp 21583 R 131072 4352 2.73 1.test |
| 54 | 11:40:16 cp 21583 R 131072 4480 6.09 1.test |
| 55 | 11:40:16 cp 21583 R 131072 5120 4.40 1.test |
| 56 | [...] |
| 57 | |
| 58 | This shows all NFS_READS that were more than 1ms. Depending on your |
| 59 | latency to your fileserver, you might need to tweak this value to |
| 60 | remove |
| 61 | |
| 62 | A threshold of 0 will trace all operations. Warning: the output will be |
| 63 | verbose, as it will include all file system cache hits. |
| 64 | |
| 65 | ./nfsslower.py 0 |
| 66 | Tracing NFS operations |
| 67 | 11:56:50 dd 21852 W 1048576 0 0.42 1.test |
| 68 | 11:56:50 dd 21852 W 1048576 1024 0.46 1.test |
| 69 | 11:56:50 dd 21852 W 1048576 2048 0.36 1.test |
| 70 | 11:56:50 cp 21854 G 0 0 0.35 1.test |
| 71 | 11:56:50 cp 21854 O 0 0 0.33 1.test |
| 72 | 11:56:50 cp 21854 G 0 0 0.00 1.test |
| 73 | 11:56:50 cp 21854 R 131072 0 0.07 1.test |
| 74 | 11:56:50 cp 21854 R 131072 128 0.02 1.test |
| 75 | 11:56:50 cp 21854 R 131072 256 0.02 1.test |
| 76 | 11:56:50 cp 21854 R 131072 384 0.02 1.test |
| 77 | 11:56:50 cp 21854 R 131072 512 0.02 1.test |
| 78 | 11:56:50 cp 21854 R 131072 640 0.02 1.test |
| 79 | 11:56:50 cp 21854 R 131072 768 0.02 1.test |
| 80 | 11:56:50 cp 21854 R 131072 896 0.02 1.test |
| 81 | 11:56:50 cp 21854 R 131072 1024 0.02 1.test |
| 82 | 11:56:50 cp 21854 R 131072 1152 0.02 1.test |
| 83 | 11:56:50 cp 21854 R 131072 1280 0.02 1.test |
| 84 | 11:56:50 cp 21854 R 131072 1408 0.02 1.test |
| 85 | 11:56:50 cp 21854 R 131072 1536 0.02 1.test |
| 86 | 11:56:50 cp 21854 R 131072 1664 0.02 1.test |
| 87 | 11:56:50 cp 21854 R 131072 1792 0.02 1.test |
| 88 | 11:56:50 cp 21854 R 131072 1920 0.02 1.test |
| 89 | 11:56:50 cp 21854 R 131072 2048 0.02 1.test |
| 90 | 11:56:50 cp 21854 R 131072 2176 0.04 1.test |
| 91 | 11:56:50 cp 21854 R 131072 2304 0.02 1.test |
| 92 | 11:56:50 cp 21854 R 131072 2432 0.03 1.test |
| 93 | 11:56:50 cp 21854 R 131072 2560 0.03 1.test |
| 94 | 11:56:50 cp 21854 R 131072 2688 0.02 1.test |
| 95 | 11:56:50 cp 21854 R 131072 2816 0.03 1.test |
| 96 | 11:56:50 cp 21854 R 131072 2944 0.02 1.test |
| 97 | 11:56:50 cp 21854 R 0 3072 0.00 1.test |
| 98 | 11:56:50 ls 21855 G 0 0 0.00 1.test |
| 99 | 11:56:50 ls 21856 G 0 0 0.36 music |
| 100 | 11:56:50 ls 21856 G 0 0 0.00 music |
| 101 | 11:56:50 ls 21856 G 0 0 0.00 test |
| 102 | 11:56:50 ls 21856 G 0 0 0.00 ff |
| 103 | 11:56:50 ls 21856 G 0 0 0.00 34.log |
| 104 | 11:56:50 ls 21856 G 0 0 0.00 vmlinuz-linux |
| 105 | 11:56:50 ls 21856 G 0 0 0.00 2.test |
| 106 | 11:56:50 ls 21856 G 0 0 0.00 rt.log |
| 107 | 11:56:50 ls 21856 G 0 0 0.00 1.lod |
| 108 | 11:56:50 ls 21856 G 0 0 0.00 COPYRIGHT.txt |
| 109 | 11:56:50 ls 21856 G 0 0 0.00 gg |
| 110 | 11:56:50 ls 21856 G 0 0 0.00 qw.log |
| 111 | 11:56:50 ls 21856 G 0 0 0.00 README.md |
| 112 | 11:56:50 ls 21856 G 0 0 0.00 1.log |
| 113 | |
| 114 | The output now includes open operations ("O"), and reads ("R") wand getattrs ("G"). |
| 115 | A cp operation |
| 116 | |
| 117 | |
| 118 | A -j option will print just the fields (parsable output, csv): |
| 119 | |
| 120 | ./nfsslower.py -j 0 |
| 121 | ENDTIME_us,TASK,PID,TYPE,BYTES,OFFSET_b,LATENCY_us,FILE |
| 122 | 87054476520,dd,22754,W,1048576,0,425,1.test |
| 123 | 87054482916,dd,22754,W,1048576,1048576,320,1.test |
| 124 | 87054488179,dd,22754,W,1048576,2097152,389,1.test |
| 125 | 87054511340,cp,22756,G,0,0,371,1.test |
| 126 | 87054511685,cp,22756,O,0,0,306,1.test |
| 127 | 87054511700,cp,22756,G,0,0,2,1.test |
| 128 | 87054512325,cp,22756,R,131072,0,56,1.test |
| 129 | 87054512432,cp,22756,R,131072,131072,22,1.test |
| 130 | 87054512520,cp,22756,R,131072,262144,32,1.test |
| 131 | 87054512600,cp,22756,R,131072,393216,21,1.test |
| 132 | 87054512678,cp,22756,R,131072,524288,21,1.test |
| 133 | 87054512803,cp,22756,R,131072,655360,56,1.test |
| 134 | |
| 135 | This may be useful for visualizing with another tool, for example, for |
| 136 | producing a scatter plot of ENDTIME vs LATENCY, to look for time-based |
| 137 | patterns. |
| 138 | |
| 139 | USAGE message: |
| 140 | |
| 141 | usage: nfsslower.py [-h] [-j] [-p PID] [min_ms] |
| 142 | |
| 143 | Trace READ, WRITE, OPEN and GETATTR NFS calls slower than a threshold,supports NFSv{3,4} |
| 144 | |
| 145 | positional arguments: |
| 146 | min_ms Minimum IO duration to trace in ms (default=10ms) |
| 147 | |
| 148 | optional arguments: |
| 149 | -h, --help show this help message and exit |
| 150 | -j, --csv just print fields: comma-separated values |
| 151 | -p PID, --pid PID Trace this pid only |
| 152 | |
| 153 | ./nfsslower # trace operations slower than 10ms |
| 154 | ./nfsslower 1 # trace operations slower than 1ms |
| 155 | ./nfsslower -j 1 # ... 1 ms, parsable output (csv) |
| 156 | ./nfsslower 0 # trace all nfs operations |
| 157 | ./nfsslower -p 121 # trace pid 121 only |
| 158 | |