Details
-
Sub-task
-
Status: Closed
-
Major
-
Resolution: Fixed
-
0.20.203.0, 0.22.0, 0.23.0
-
None
-
Reviewed
-
block report, performance, startup
Description
Consider a large cluster that takes 40 minutes to start up. The datanodes compete to register and send their Initial Block Reports (IBRs) as fast as they can after startup (subject to a small sub-two-minute random delay, which isn't relevant to this discussion).
As each datanode succeeds in sending its IBR, it schedules the starting time for its regular cycle of reports, every hour (or other configured value of dfs.blockreport.intervalMsec). In order to spread the reports evenly across the block report interval, each datanode picks a random fraction of that interval, for the starting point of its regular report cycle. For example, if a particular datanode ends up randomly selecting 18 minutes after the hour, then that datanode will send a Block Report at 18 minutes after the hour every hour as long as it remains up. Other datanodes will start their cycles at other randomly selected times. This code is in DataNode.blockReport() and DataNode.scheduleBlockReport().
The "second Block Report" (2BR), is the start of these hourly reports. The problem is that some of these 2BRs get scheduled sooner rather than later, and actually occur within the startup period. For example, if the cluster takes 40 minutes (2/3 of an hour) to start up, then out of the datanodes that succeed in sending their IBRs during the first 10 minutes, between 1/2 and 2/3 of them will send their 2BR before the 40-minute startup time has completed!
2BRs sent within the startup time actually compete with the remaining IBRs, and thereby slow down the overall startup process. This can be seen in the following data, which shows the startup process for a 3700-node cluster that took about 17 minutes to finish startup:
time starts sum regs sum IBR sum 2nd_BR sum total_BRs/min 0 1299799498 3042 3042 1969 1969 151 151 0 151 1 1299799558 665 3707 1470 3439 248 399 0 248 2 1299799618 3707 224 3663 270 669 0 270 3 1299799678 3707 14 3677 261 930 3 3 264 4 1299799738 3707 23 3700 288 1218 1 4 289 5 1299799798 3707 7 3707 258 1476 3 7 261 6 1299799858 3707 3707 317 1793 4 11 321 7 1299799918 3707 3707 292 2085 6 17 298 8 1299799978 3707 3707 292 2377 8 25 300 9 1299800038 3707 3707 272 2649 25 272 10 1299800098 3707 3707 280 2929 15 40 295 11 1299800158 3707 3707 223 3152 14 54 237 12 1299800218 3707 3707 143 3295 54 143 13 1299800278 3707 3707 141 3436 20 74 161 14 1299800338 3707 3707 195 3631 78 152 273 15 1299800398 3707 3707 51 3682 209 361 260 16 1299800458 3707 3707 25 3707 369 730 394 17 1299800518 3707 3707 3707 166 896 166 18 1299800578 3707 3707 3707 72 968 72 19 1299800638 3707 3707 3707 67 1035 67 20 1299800698 3707 3707 3707 75 1110 75 21 1299800758 3707 3707 3707 71 1181 71 22 1299800818 3707 3707 3707 67 1248 67 23 1299800878 3707 3707 3707 62 1310 62 24 1299800938 3707 3707 3707 56 1366 56 25 1299800998 3707 3707 3707 60 1426 60
This data was harvested from the startup logs of all the datanodes, and correlated into one-minute buckets. Each row of the table represents the progress during one elapsed minute of clock time. It seems that every cluster startup is different, but this one showed the effect fairly well.
The "starts" column shows that all the nodes started up within the first 2 minutes, and the "regs" column shows that all succeeded in registering by minute 6. The IBR column shows a sustained rate of Initial Block Report processing of 250-300/minute for the first 10 minutes.
The question is why, during minutes 11 through 16, the rate of IBR processing slowed down. Why didn't the startup just finish? In the "2nd_BR" column, we see the rate of 2BRs ramping up as more datanodes complete their IBRs. As the rate increases, they become more effective at competing with the IBRs, and slow down the IBR processing even more. After the IBRs finally finish in minute 16, the rate of 2BRs settles down to a steady ~60-70/minute.
In order to decrease competition for locks and other resources, to speed up IBR processing during startup, we propose to delay 2BRs until later into the cycle.