Abstract: Reducing file write latency includes receiving incoming data, from a data source, for storage in a file and a target storage location for the incoming data, and determining whether the target storage location corresponds to a cache entry. Based on at least the target storage location not corresponding to a cache entry, the incoming data is written to a block pre-allocated for cache misses and the writing of the incoming data to the pre-allocated block is journaled. The writing of the incoming data is acknowledged to the data source. A process executing in parallel with the above commits the incoming data in the pre-allocated block with the file. Using this parallel process to commit the incoming data in the file removes high-latency operations (e.g., reading pointer blocks from the storage media) from a critical input/output path and results in more rapid write acknowledgement.

Abstract: Techniques are disclosed for dynamically managing a cache. Certain techniques include clustering I/O requests into a plurality of clusters by a machine-learning clustering algorithm that collects the I/O requests into clusters of similar I/O requests based on properties of the I/O requests. Further, certain techniques include identifying, for a received I/O request, a cluster stored in the cache. Certain techniques further include loading a set of blocks of the identified cluster into the cache.

Abstract: Reducing file write latency includes receiving incoming data, from a data source, for storage in a file and a target storage location for the incoming data, and determining whether the target storage location corresponds to a cache entry. Based on at least the target storage location not corresponding to a cache entry, the incoming data is written to a block pre-allocated for cache misses and the writing of the incoming data to the pre-allocated block is journaled. The writing of the incoming data is acknowledged to the data source. A process executing in parallel with the above commits the incoming data in the pre-allocated block with the file. Using this parallel process to commit the incoming data in the file removes high-latency operations (e.g., reading pointer blocks from the storage media) from a critical input/output path and results in more rapid write acknowledgement.

Abstract: The disclosure provides for fault tolerant parallel journaling that speeds up both input/output (I/O) operations and recovery operations. Journal entry writing may occur in parallel with data writing operations. Even if a crash occurs during a data writing operation for which the journal entry has been written, the recovery operation will correctly determine that the journal entry is not valid. Additionally, recovery operations may need to validate fewer journal entries, and yet possibly retain more valid data. Examples include: for each of a plurality of journal entries: receiving incoming data; determining a signature for the incoming data; generating the journal entry for the incoming data; writing the signature in the journal entry; and writing the journal entry and the incoming data to a storage media; and based at least on writing data to the storage media, updating an awaiting index in a journal header.

Abstract: The disclosure supports both trickle and burst input/output (I/O) admission rates in journaling file systems. Examples include receiving incoming data; based at least on receiving the incoming data, generating metadata for a journal entry; adding the metadata to an active metadata batch; issuing a data write to write the incoming data to a storage medium; monitoring for a first trigger comprising determining that a data write for an entry in the active metadata batch is complete; based at least on the first trigger, closing the active metadata batch; and issuing a journal write to write entries of the active metadata batch to the storage medium. A second trigger comprises determining that a batch open time exceeds a selected percentage of a moving average of data write durations. A third trigger comprises determining that a batch counter exceeds a count threshold. These triggers work together to reduce I/O latencies.

Abstract: Techniques are disclosed for dynamically managing a cache. Certain techniques include clustering I/O requests into a plurality of clusters by a machine-learning clustering algorithm that collects the I/O requests into clusters of similar I/O requests based on properties of the I/O requests. Further, certain techniques include identifying, for a received I/O request, a cluster stored in the cache. Certain techniques further include loading a set of blocks of the identified cluster into the cache.

Abstract: The disclosure supports both trickle and burst input/output (I/O) admission rates in journaling file systems. Examples include receiving incoming data; based at least on receiving the incoming data, generating metadata for a journal entry; adding the metadata to an active metadata batch; issuing a data write to write the incoming data to a storage medium; monitoring for a first trigger comprising determining that a data write for an entry in the active metadata batch is complete; based at least on the first trigger, closing the active metadata batch; and issuing a journal write to write entries of the active metadata batch to the storage medium. A second trigger comprises determining that a batch open time exceeds a selected percentage of a moving average of data write durations. A third trigger comprises determining that a batch counter exceeds a count threshold. These triggers work together to reduce I/O latencies.

Abstract: The disclosure provides for fault tolerant parallel journaling that speeds up both input/output (I/O) operations and recovery operations. Journal entry writing may occur in parallel with data writing operations. Even if a crash occurs during a data writing operation for which the journal entry has been written, the recovery operation will correctly determine that the journal entry is not valid. Additionally, recovery operations may need to validate fewer journal entries, and yet possibly retain more valid data. Examples include: for each of a plurality of journal entries: receiving incoming data; determining a signature for the incoming data; generating the journal entry for the incoming data; writing the signature in the journal entry; and writing the journal entry and the incoming data to a storage media; and based at least on writing data to the storage media, updating an awaiting index in a journal header.