In one embodiment, a parallel (e.g., tiered) logging technique is provided to deliver low latency acknowledgements of input/output (I/O) requests, such as write requests, while avoiding loss of data. Write data may be stored (copied) as a log in a portion of a dynamic random access memory and a non-volatile random access memory (NVRAM). The NVRAM may be configured as, e.g., a persistent write-back cache of the node, while parameters of the request may be stored in another portion of the NVRAM configured as the log (NVLog). The write data may be organized into separate variable length blocks or extents and “written back” out-of-order from the write-back cache to storage devices, such as SSDs, e.g., organized into a data container (intended destination of the write request). The write data may be preserved in the NVlog until each extent is safely stored on SSD.