H2O双节点集群运行XGBoost模型时的资源共享问题排查求助
H2O双节点集群资源分配异常排查求助
使用H2O-3.36.1.1版本搭建双节点集群,基于XGBoost建模过程中出现以下异常:
- 集群无法共享内存资源:Master A服务器内存占用极高,Master B服务器内存占用极低
- 模型运行阶段,Master A的H2O日志持续更新,但Master B的日志仅记录集群创建相关内容,无后续更新
- 运行过程中,Master A上的H2O Jar偶尔会因内存占用过高崩溃
已确认集群创建成功(日志记录集群详情),且Master A与Master B双向curl连通性正常,集群整体状态显示为locked, healthy。
集群详细信息
- H2O_cluster_uptime: 15 mins 14 secs
- H2O_cluster_timezone: Asia/Colombo
- H2O_data_parsing_timezone: UTC
- H2O_cluster_version: 3.36.1.1
- H2O_cluster_version_age: 11 months and 28 days !!!
- H2O_cluster_name: XXXXXX
- H2O_cluster_total_nodes: 2
- H2O_cluster_free_memory: 43.36 Gb
- H2O_cluster_total_cores: 30
- H2O_cluster_allowed_cores: 30
- H2O_cluster_status: locked, healthy
- H2O_connection_url: http://localhost:54321
- H2O_connection_proxy: {"http": null, "https": null}
- H2O_internal_security: False
- Python_version: 3.7.11 final
需要排查的问题
- 为何两台服务器在模型运行阶段无法共享资源?
- 为何Master B的H2O日志无后续更新?
- 为何Master A上的H2O Jar会因内存占用过高崩溃?
Master A日志
main INFO water.default: Open H2O Flow in your web browser: http://xxx.xxx.xxx.xx:54321 main INFO water.default: FJ-126-15 INFO water.default: Cloud of size 2 formed [master01.user.com/xxx.xxx.xxx.xx:54321, master02.user.com/xxx.xxx.xxx.xx:54321] 058452-166 INFO water.default: GET /3/Metadata/schemas/CloudV3, parms: {} 058452-166 INFO water.default: Locking cloud to new members, because water.api.schemas3.MetadataV3 4058452-14 INFO water.default: GET /3/Metadata/schemas/H2OErrorV3, parms: {} 4058452-15 INFO water.default: GET /3/Metadata/schemas/H2OModelBuilderErrorV3, parms: {} 4058452-18 INFO water.default: POST /4/sessions, parms: {} 4058452-16 INFO water.default: POST /99/Rapids, parms: {ast=(setTimeZone "UTC"), session_id=_sid_a391} 4058452-13 INFO water.default: DELETE /3/DKV, parms: {} 4058452-13 INFO water.default: Removing all objects 4058452-13 INFO water.default: Finished removing objects 4058452-12 INFO water.default: DELETE /3/DKV, parms: {} 4058452-12 INFO water.default: Removing all objects 4058452-12 INFO water.default: Finished removing objects 058452-170 INFO water.default: DELETE /3/DKV, parms: {} 058452-170 INFO water.default: Removing all objects 058452-170 INFO water.default: Finished removing objects 4058452-14 INFO water.default: GET /3/Metadata/schemas/CloudV3, parms: {} 058452-169 INFO water.default: GET /3/Metadata/schemas/H2OErrorV3, parms: {} 058452-166 INFO water.default: GET /3/Metadata/schemas/H2OModelBuilderErrorV3, parms: {} 4058452-19 INFO water.default: POST /4/sessions, parms: {} 4058452-18 INFO water.default: POST /99/Rapids, parms: {ast=(setTimeZone "UTC"), session_id=_sid_bfac} 058452-170 INFO water.default: Reading byte InputStream into Frame: 058452-170 INFO water.default: frameKey: upload_bbcd4f6aeb3c1095e63f66a89cdd4756 058452-170 INFO water.default: totalChunks: 2 058452-170 INFO water.default: totalBytes: 4404663 058452-170 INFO water.default: Success. 058452-167 INFO water.default: POST /3/ParseSetup, parms: {single_quotes=False, source_frames=["upload_bbcd4f6aeb3c1095e63f66a89cdd4756"], check_header=0} 058452-169 INFO water.default: Total file size: 4.2 MB 058452-169 INFO water.default: Parse chunk size 4194304 FJ-1-15 INFO water.default: Parse result for Key_Frame__upload_bbcd4f6aeb3c1095e63f66a89cdd4756.hex (2023 rows, 436 columns): FJ-1-15 INFO water.default: ColV2 type min max mean sigma NAs constant cardinality FJ-1-15 INFO water.default: COL1: factor 011022232 YA9854024 1334 FJ-1-15 INFO water.default: COL2: numeric 2019.00 2020.00 2019.70 0.457960 FJ-1-15 INFO water.default: COL3: numeric 1.00000 12.0000 6.07860 2.82287 FJ-1-15 INFO water.default: COL4: factor |00011000813 |09988000074 1334 FJ-1-15 INFO water.default: COL5: factor CUST NAME CUSTOMER 2 FJ-1-15 INFO water.default: COL6: numeric 1.14005e+08 4.10024e+08 2.96146e+08 4.57328e+07 FJ-1-15 INFO water.default: COL7: numeric 10000.0 30000.0 28294.6 5573.93 3 FJ-1-15 INFO water.default: COL8: factor USD 4 FJ-1-15 INFO water.default: COL9: factor 927 RM17 20 FJ-1-15 INFO water.default: COL10: factor NO YES 2 FJ-1-15 INFO water.default: Additional column information only sent to log file... FJ-1-15 INFO water.default: COL11: numeric -1.00000 175.250 1.07602 5.07740 FJ-1-15 INFO water.default: COL12: numeric -1.00000 97.2262 0.447662 3.19167 FJ-1-15 INFO water.default: COL13: numeric -1.00000 124.206 1.03933 3.94221 FJ-1-15 INFO water.default: response_class: factor 1A to_be_filled 5 FJ-1-15 INFO water.default: response_class_5: factor 1B 1B1 2 FJ-1-15 INFO water.default: response_class_4: factor 1A NON_PERFORME 4 FJ-1-15 INFO water.default: response_class_3: factor 1A NON_PERFORME 4 FJ-1-15 INFO water.default: response_class_2: factor 1A NON_PERFORME 4 FJ-1-15 INFO water.default: response_class_1: factor 1A NON_PERFORME 4 FJ-1-15 INFO water.default: subset: factor test train 2 FJ-1-15 INFO water.default: Chunk compression summary: FJ-1-15 INFO water.default: Chunk Type Chunk Name Count Count Percentage Size Size Percentage FJ-1-15 INFO water.default: C0L Constant long 74 8.486 % 5.8 KB 0.207 % FJ-1-15 INFO water.default: CBS Binary 19 2.179 % 4.4 KB 0.159 % FJ-1-15 INFO water.default: CXI Sparse Integers 80 9.174 % 25.0 KB 0.897 % FJ-1-15 INFO water.default: CXF Sparse Reals 50 5.734 % 48.9 KB 1.753 % FJ-1-15 INFO water.default: C1 1-Byte Integers 7 0.803 % 11.8 KB 0.423 % FJ-1-15 INFO water.default: C1N 1-Byte Integers (w/o NAs) 92 10.550 % 104.0 KB 3.731 % FJ-1-15 INFO water.default: C1S 1-Byte Fractions 142 16.284 % 118.4 KB 4.245 % FJ-1-15 INFO water.default: C2 2-Byte Integers 72 8.257 % 231.7 KB 8.309 % FJ-1-15 INFO water.default: C2S 2-Byte Fractions 18 2.064 % 22.9 KB 0.822 % FJ-1-15 INFO water.default: C4 4-Byte Integers 50 5.734 % 109.1 KB 3.913 % FJ-1-15 INFO water.default: C4S 4-Byte Fractions 127 14.564 % 360.5 KB 12.925 % FJ-1-15 INFO water.default: C8 8-byte Integers 1 0.115 % 15.0 KB 0.539 % FJ-1-15 INFO water.default: CUD Unique Reals 5 0.573 % 13.2 KB 0.472 % FJ-1-15 INFO water.default: C8D 64-bit Reals 135 15.482 % 1.7 MB 61.606 % FJ-1-15 INFO water.default: Frame distribution summary: FJ-1-15 INFO water.default: Size Number of Rows Number of Chunks per Column Number of Chunks
Master B日志
main INFO water.default: H2O started in 4906ms main INFO water.default: main INFO water.default: Open H2O Flow in your web browser: http://xxx.xxx.xxx.xx:54321 main INFO water.default: FJ-126-15 INFO water.default: Cloud of size 2 formed [master01.user.com/xxx.xxx.xxx.xx:54321, master02.user.com/xxx.xxx.xxx.xx:54321] FJ-123-15 INFO water.default: Locking cloud to new members, because Class Id=56 FJ-2-15 INFO water.default: Key upload_bbcd4f6aeb3c1095e63f66a89cdd4756 will be parsed using method DistributedParse. FJ-2-21 INFO water.default: Key upload_902bcdd31a4aea9f65690f1bc6074886 will be parsed using method DistributedParse.
内容的提问来源于stack exchange,提问作者Nalinda Perera
相关产品推荐
相关产品推荐

