DeepSpeed: DeepSpeed is slower than FSDP

Describe the bug I am still familiarizing with DeepSpeed so here is a n00b question. I wrapped my model with DeepSpeed and seeing good ZeRO2 performance. However, when I switch to ZeRO3, the all gathers are not overlapping and they are very fragmented even though the default params look good. How can I learn more about why all gathers are so fragmented and how to make them less granular?

DeepSpeedEngine configuration:
activation_checkpointing_config  {
    "partition_activations": false, 
    "contiguous_memory_optimization": false, 
    "cpu_checkpointing": false, 
    "number_checkpoints": null, 
    "synchronize_checkpoint_boundary": false, 
    "profile": false
}
aio_config ................... {'block_size': 1048576, 'queue_depth': 8, 'thread_count': 1, 'single_submit': False, 'overlap_events': True}
amp_enabled .................. False
amp_params ................... False
autotuning_config ............ {
    "enabled": false, 
    "start_step": null, 
    "end_step": null, 
    "metric_path": null, 
    "arg_mappings": null, 
    "metric": "throughput", 
    "model_info": null, 
    "results_dir": "autotuning_results", 
    "exps_dir": "autotuning_exps", 
    "overwrite": true, 
    "fast": true, 
    "start_profile_step": 3, 
    "end_profile_step": 5, 
    "tuner_type": "gridsearch", 
    "tuner_early_stopping": 5, 
    "tuner_num_trials": 50, 
    "model_info_path": null, 
    "mp_size": 1, 
    "max_train_batch_size": null, 
    "min_train_batch_size": 1, 
    "max_train_micro_batch_size_per_gpu": 1.024000e+03, 
    "min_train_micro_batch_size_per_gpu": 1, 
    "num_tuning_micro_batch_sizes": 3
}
bfloat16_enabled ............. True
checkpoint_parallel_write_pipeline  False
checkpoint_tag_validation_enabled  True
checkpoint_tag_validation_fail  False
comms_config ................. <deepspeed.comm.config.DeepSpeedCommsConfig object at 0x7f09382c8a90>
communication_data_type ...... None
compression_config ........... {'weight_quantization': {'shared_parameters': {'enabled': False, 'quantizer_kernel': False, 'schedule_offset': 0, 'quantize_groups': 1, 'quantize_verbose': False, 'quantization_type': 'symmetric', 'quantize_weight_in_forward': False, 'rounding': 'nearest', 'fp16_mixed_quantize': False, 'quantize_change_ratio': 0.001}, 'different_groups': {}}, 'activation_quantization': {'shared_parameters': {'enabled': False, 'quantization_type': 'symmetric', 'range_calibration': 'dynamic', 'schedule_offset': 1000}, 'different_groups': {}}, 'sparse_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'row_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'head_pruning': {'shared_parameters': {'enabled': False, 'method': 'topk', 'schedule_offset': 1000}, 'different_groups': {}}, 'channel_pruning': {'shared_parameters': {'enabled': False, 'method': 'l1', 'schedule_offset': 1000}, 'different_groups': {}}, 'layer_reduction': {'enabled': False}}
curriculum_enabled_legacy .... False
curriculum_params_legacy ..... False
data_efficiency_config ....... {'enabled': False, 'seed': 1234, 'data_sampling': {'enabled': False, 'num_epochs': 1000, 'num_workers': 0, 'curriculum_learning': {'enabled': False}}, 'data_routing': {'enabled': False, 'random_ltd': {'enabled': False, 'layer_token_lr_schedule': {'enabled': False}}}}
data_efficiency_enabled ...... False
dataloader_drop_last ......... False
disable_allgather ............ False
dump_state ................... False
dynamic_loss_scale_args ...... None
eigenvalue_enabled ........... False
eigenvalue_gas_boundary_resolution  1
eigenvalue_layer_name ........ bert.encoder.layer
eigenvalue_layer_num ......... 0
eigenvalue_max_iter .......... 100
eigenvalue_stability ......... 1e-06
eigenvalue_tol ............... 0.01
eigenvalue_verbose ........... False
elasticity_enabled ........... False
flops_profiler_config ........ {
    "enabled": false, 
    "recompute_fwd_factor": 0.0, 
    "profile_step": 1, 
    "module_depth": -1, 
    "top_modules": 1, 
    "detailed": true, 
    "output_file": null
}
fp16_auto_cast ............... None
fp16_enabled ................. False
fp16_master_weights_and_gradients  False
global_rank .................. 0
grad_accum_dtype ............. None
gradient_accumulation_steps .. 4
gradient_clipping ............ 1.0
gradient_predivide_factor .... 1.0
graph_harvesting ............. False
hybrid_engine ................ enabled=False max_out_tokens=512 inference_tp_size=1 release_inference_cache=False pin_parameters=True tp_gather_partition_size=8
initial_dynamic_scale ........ 1
load_universal_checkpoint .... False
loss_scale ................... 1.0
memory_breakdown ............. False
mics_hierarchial_params_gather  False
mics_shard_size .............. -1
monitor_config ............... tensorboard=TensorBoardConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') wandb=WandbConfig(enabled=False, group=None, team=None, project='deepspeed') csv_monitor=CSVConfig(enabled=False, output_path='', job_name='DeepSpeedJobName') enabled=False
nebula_config ................ {
    "enabled": false, 
    "persistent_storage_path": null, 
    "persistent_time_interval": 100, 
    "num_of_version_in_retention": 2, 
    "enable_nebula_load": true, 
    "load_path": null
}
optimizer_legacy_fusion ...... False
optimizer_name ............... None
optimizer_params ............. None
pipeline ..................... {'stages': 'auto', 'partition': 'best', 'seed_layers': False, 'activation_checkpoint_interval': 0, 'pipe_partitioned': True, 'grad_partitioned': True}
pld_enabled .................. False
pld_params ................... False
prescale_gradients ........... False
scheduler_name ............... None
scheduler_params ............. None
seq_parallel_communication_data_type  torch.float32
sparse_attention ............. None
sparse_gradients_enabled ..... False
steps_per_print .............. 1
train_batch_size ............. 1024
train_micro_batch_size_per_gpu  8
use_data_before_expert_parallel_  False
use_node_local_storage ....... False
wall_clock_breakdown ......... False
weight_quantization_config ... None
world_size ................... 32
zero_allow_untested_optimizer  False
zero_config
..................
stage=3
contiguous_gradients=True
reduce_scatter=True
reduce_bucket_size=500000000
use_multi_rank_bucket_allreduce=True
allgather_partitions=True
allgather_bucket_size=500000000
overlap_comm=True
load_from_fp32_weights=True
elastic_checkpoint=False
offload_param=None
offload_optimizer=None
sub_group_size=1,000,000,000
cpu_offload_param=None
cpu_offload_use_pin_memory=None
cpu_offload=None
prefetch_bucket_size=50,000,000
param_persistence_threshold=100,000
model_persistence_threshold=sys.maxsize
max_live_parameters=1,000,000,000
max_reuse_distance=1,000,000,000
gather_16bit_weights_on_model_save=False
stage3_gather_fp16_weights_on_model_save=False
ignore_unused_parameters=True
legacy_stage1=False
round_robin_gradients=False
zero_hpz_partition_size=1
zero_quantized_weights=False
zero_quantized_nontrainable_weights=False
zero_quantized_gradients=False
mics_shard_size=-1
mics_hierarchical_params_gather=False
memory_efficient_linear=True
pipeline_loading_checkpoint=False
override_module_apply=True