Wrong balance between Aerospike instances in cluster - load-balancing

Question

I have an application with a high load for batch read operations. My Aerospike cluster (v 3.7.2) has 14 servers, each one with 7GB RAM and 2 CPUs in Google Cloud.
By looking at Google Cloud Monitoring Graphs, I noticed a very unbalanced load between servers: some servers have almost 100% CPU load, while others have less than 50% (image below). Even after hours of operation, the cluster unbalanced pattern doesn't change.
Is there any configuration that I could change to make this cluster more homogeneous? How to optimize node balancing?
Edit 1
All servers in the cluster have the same identical aerospike.conf file:
Aerospike database configuration file.
service {
user root
group root
paxos-single-replica-limit 1 # Number of nodes where the replica count is automatically reduced to 1.
paxos-recovery-policy auto-reset-master
pidfile /var/run/aerospike/asd.pid
service-threads 32
transaction-queues 32
transaction-threads-per-queue 32
batch-index-threads 32
proto-fd-max 15000
batch-max-requests 200000
}
logging {
# Log file must be an absolute path.
file /var/log/aerospike/aerospike.log {
context any info
}
}
network {
service {
#address any
port 3000
}
heartbeat {
mode mesh
mesh-seed-address-port 10.240.0.6 3002
mesh-seed-address-port 10.240.0.5 3002
port 3002
interval 150
timeout 20
}
fabric {
port 3001
}
info {
port 3003
}
}
namespace test {
replication-factor 3
memory-size 5G
default-ttl 0 # 30 days, use 0 to never expire/evict.
ldt-enabled true
storage-engine device {
file /data/aerospike.dat
write-block-size 1M
filesize 180G
}
}
Edit 2:
$ asinfo
1 : node
BB90600F00A0142
2 : statistics
cluster_size=14;cluster_key=E3C3672DCDD7F51;cluster_integrity=true;objects=3739898;sub-records=0;total-bytes-disk=193273528320;used-bytes-disk=26018492544;free-pct-disk=86;total-bytes-memory=5368709120;used-bytes-memory=239353472;data-used-bytes-memory=0;index-used-bytes-memory=239353472;sindex-used-bytes-memory=0;free-pct-memory=95;stat_read_reqs=2881465329;stat_read_reqs_xdr=0;stat_read_success=2878457632;stat_read_errs_notfound=3007093;stat_read_errs_other=0;stat_write_reqs=551398;stat_write_reqs_xdr=0;stat_write_success=549522;stat_write_errs=90;stat_xdr_pipe_writes=0;stat_xdr_pipe_miss=0;stat_delete_success=4;stat_rw_timeout=1862;udf_read_reqs=0;udf_read_success=0;udf_read_errs_other=0;udf_write_reqs=0;udf_write_success=0;udf_write_err_others=0;udf_delete_reqs=0;udf_delete_success=0;udf_delete_err_others=0;udf_lua_errs=0;udf_scan_rec_reqs=0;udf_query_rec_reqs=0;udf_replica_writes=0;stat_proxy_reqs=7021;stat_proxy_reqs_xdr=0;stat_proxy_success=2121;stat_proxy_errs=4739;stat_ldt_proxy=0;stat_cluster_key_err_ack_dup_trans_reenqueue=607;stat_expired_objects=0;stat_evicted_objects=0;stat_deleted_set_objects=0;stat_evicted_objects_time=0;stat_zero_bin_records=0;stat_nsup_deletes_not_shipped=0;stat_compressed_pkts_received=0;err_tsvc_requests=110;err_tsvc_requests_timeout=0;err_out_of_space=0;err_duplicate_proxy_request=0;err_rw_request_not_found=17;err_rw_pending_limit=19;err_rw_cant_put_unique=0;geo_region_query_count=0;geo_region_query_cells=0;geo_region_query_points=0;geo_region_query_falsepos=0;fabric_msgs_sent=58002818;fabric_msgs_rcvd=57998870;paxos_principal=BB92B00F00A0142;migrate_msgs_sent=55749290;migrate_msgs_recv=55759692;migrate_progress_send=0;migrate_progress_recv=0;migrate_num_incoming_accepted=7228;migrate_num_incoming_refused=0;queue=0;transactions=101978550;reaped_fds=6;scans_active=0;basic_scans_succeeded=0;basic_scans_failed=0;aggr_scans_succeeded=0;aggr_scans_failed=0;udf_bg_scans_succeeded=0;udf_bg_scans_failed=0;batch_index_initiate=40457778;batch_index_queue=0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0,0:0;batch_index_complete=40456708;batch_index_timeout=1037;batch_index_errors=33;batch_index_unused_buffers=256;batch_index_huge_buffers=217168717;batch_index_created_buffers=217583519;batch_index_destroyed_buffers=217583263;batch_initiate=0;batch_queue=0;batch_tree_count=0;batch_timeout=0;batch_errors=0;info_queue=0;delete_queue=0;proxy_in_progress=0;proxy_initiate=7021;proxy_action=5519;proxy_retry=0;proxy_retry_q_full=0;proxy_unproxy=0;proxy_retry_same_dest=0;proxy_retry_new_dest=0;write_master=551089;write_prole=1055431;read_dup_prole=14232;rw_err_dup_internal=0;rw_err_dup_cluster_key=1814;rw_err_dup_send=0;rw_err_write_internal=0;rw_err_write_cluster_key=0;rw_err_write_send=0;rw_err_ack_internal=0;rw_err_ack_nomatch=1767;rw_err_ack_badnode=0;client_connections=366;waiting_transactions=0;tree_count=0;record_refs=3739898;record_locks=0;migrate_tx_objs=0;migrate_rx_objs=0;ongoing_write_reqs=0;err_storage_queue_full=0;partition_actual=296;partition_replica=572;partition_desync=0;partition_absent=3228;partition_zombie=0;partition_object_count=3739898;partition_ref_count=4096;system_free_mem_pct=61;sindex_ucgarbage_found=0;sindex_gc_locktimedout=0;sindex_gc_inactivity_dur=0;sindex_gc_activity_dur=0;sindex_gc_list_creation_time=0;sindex_gc_list_deletion_time=0;sindex_gc_objects_validated=0;sindex_gc_garbage_found=0;sindex_gc_garbage_cleaned=0;system_swapping=false;err_replica_null_node=0;err_replica_non_null_node=0;err_sync_copy_null_master=0;storage_defrag_corrupt_record=0;err_write_fail_prole_unknown=0;err_write_fail_prole_generation=0;err_write_fail_unknown=0;err_write_fail_key_exists=0;err_write_fail_generation=0;err_write_fail_generation_xdr=0;err_write_fail_bin_exists=0;err_write_fail_parameter=0;err_write_fail_incompatible_type=0;err_write_fail_noxdr=0;err_write_fail_prole_delete=0;err_write_fail_not_found=0;err_write_fail_key_mismatch=0;err_write_fail_record_too_big=90;err_write_fail_bin_name=0;err_write_fail_bin_not_found=0;err_write_fail_forbidden=0;stat_duplicate_operation=53184;uptime=1001388;stat_write_errs_notfound=0;stat_write_errs_other=90;heartbeat_received_self=0;heartbeat_received_foreign=145137042;query_reqs=0;query_success=0;query_fail=0;query_abort=0;query_avg_rec_count=0;query_short_running=0;query_long_running=0;query_short_queue_full=0;query_long_queue_full=0;query_short_reqs=0;query_long_reqs=0;query_agg=0;query_agg_success=0;query_agg_err=0;query_agg_abort=0;query_agg_avg_rec_count=0;query_lookups=0;query_lookup_success=0;query_lookup_err=0;query_lookup_abort=0;query_lookup_avg_rec_count=0
3 : features
cdt-list;pipelining;geo;float;batch-index;replicas-all;replicas-master;replicas-prole;udf
4 : cluster-generation
61
5 : partition-generation
11811
6 : edition
Aerospike Community Edition
7 : version
Aerospike Community Edition build 3.7.2
8 : build
3.7.2
9 : services
10.0.3.1:3000;10.240.0.14:3000;10.0.3.1:3000;10.240.0.27:3000;10.0.3.1:3000;10.240.0.5:3000;10.0.3.1:3000;10.240.0.43:3000;10.0.3.1:3000;10.240.0.30:3000;10.0.3.1:3000;10.240.0.18:3000;10.0.3.1:3000;10.240.0.42:3000;10.0.3.1:3000;10.240.0.33:3000;10.0.3.1:3000;10.240.0.24:3000;10.0.3.1:3000;10.240.0.37:3000;10.0.3.1:3000;10.240.0.41:3000;10.0.3.1:3000;10.240.0.13:3000;10.0.3.1:3000;10.240.0.23:3000
10 : services-alumni
10.0.3.1:3000;10.240.0.42:3000;10.0.3.1:3000;10.240.0.5:3000;10.0.3.1:3000;10.240.0.13:3000;10.0.3.1:3000;10.240.0.14:3000;10.0.3.1:3000;10.240.0.18:3000;10.0.3.1:3000;10.240.0.23:3000;10.0.3.1:3000;10.240.0.24:3000;10.0.3.1:3000;10.240.0.27:3000;10.0.3.1:3000;10.240.0.30:3000;10.0.3.1:3000;10.240.0.37:3000;10.0.3.1:3000;10.240.0.43:3000;10.0.3.1:3000;10.240.0.33:3000;10.0.3.1:3000;10.240.0.41:3000

Answer 1

I have a few comments about your configuration. First, transaction-threads-per-queue should be set to 3 or 4 (don't set it to the number of cores).
The second has to do with your batch-read tuning. You're using the (default) batch-index protocol, and the config params you'll need to tune for batch-read performance are:
You have batch-max-requests set very high. This is probably affecting both your CPU load and your memory consumption. It's enough that there's a slight imbalance in the number of keys you're accessing per-node, and that will reflect in the graphs you've shown. At least, this is possibly the issue. It's better that you iterate over smaller batches than try to fetch 200K records per-node at a time.
batch-index-threads – by default its value is 4, and you set it to 32 (of a max of 64). You should do this incrementally by running the same test and benchmarking the performance. On each iteration adjust higher, then down if it's decreased in performance. For example: test with 32, +8 = 40 , +8 = 48, -4 = 44. There's no easy rule-of-thumb for the setting, you'll need to tune through iterations on the hardware you'll be using, and monitor the performance.
batch-max-buffer-per-queue – this is more directly linked to the number of concurrent batch-read operations the node can support. Each batch-read request will consume at least one buffer (more if the data cannot fit in 128K). If you do not have enough of these allocated to support the number of concurrent batch-read requests you will get exceptions with error code 152 BATCH_QUEUES_FULL . Track and log such events clearly, because it means you need to raise this value. Note that this is the number of buffers per-queue. Each batch response worker thread has its own queue, so you'll have batch-index-threads x batch-max-buffer-per-queue buffers, each taking 128K of RAM. The batch-max-unused-buffers caps the memory usage of all these buffers combined, destroying unused buffers until their number is reduced. There's an overhead to allocating and destroying these buffers, so you do not want to set it too low compared to the total. Your current cost is 32 x 256 x 128KB = 1GB.
Finally, you're storing your data on a filesystem. That's fine for development instances, but not recommended for production. In GCE you can provision either a SATA SSD or an NVMe SSD for your data storage, and those should be initialized, and used as block devices. Take a look at the GCE recommendations for more details. I suspect you have warnings in your log about the device not keeping up.
It's likely that one of your nodes is an outlier with regards to the number of partitions it has (and therefore number of objects). You can confirm it with asadm -e 'asinfo -v "objects"'. If that's the case, you can terminate that node, and bring up a new one. This will force the partitions to be redistributed. This does trigger a migration, which takes quite longer in the CE server than in the EE one.

Answer 2

For anyone interested, Aerospike Enterpirse 4.3 introduced 'uniform-balance' which homogeneously balances data partitions. Read more here: https://www.aerospike.com/blog/aerospike-4-3-all-flash-uniform-balance/