Merge branch 'main' into dynamic-default-part-size

Author: TingDaoK

include/aws/s3/s3_client.h

@@ -317,29 +317,26 @@ enum aws_s3_recv_file_options {
 
 /**
  * WARNING: experimental/unstable:
- * Controls how client performance file I/O operations. Only applies to the file based
- * workload.
+ * Controls how client performance file I/O operations. Only applies to the file based workload.
  **/
 struct aws_s3_file_io_options {
     /**
      * Skip buffering the part in memory before sending the request.
-     * If set, set the `disk_throughput_gbps` to be reasonable align with the available disk throughput.
-     * Otherwise, the transfer may fail with connection starvation.
      *
-     * Default to false.
+     * Default to false on small objects, and true when the object size exceed a certain threshold
+     *`g_streaming_object_size_threshold`.
      **/
     bool should_stream;
 
     /**
      * The estimated disk throughput. Only be applied when `streaming_upload` is true.
      * in gigabits per second (Gbps).
      *
-     * When doing upload with streaming, it's important to set the disk throughput to prevent the connection starvation.
      * Notes: There are possibilities that cannot reach the all available disk throughput:
      * 1. Disk is busy with other applications
      * 2. OS Cache may cap the throughput, use `direct_io` to get around this.
      *
-     * Note: When `streaming_upload` is true, this default to 10 Gbps.
+     * Default to throughput_target_gbps.
      **/
     double disk_throughput_gbps;
 

source/s3_meta_request.c

@@ -194,7 +194,7 @@ int aws_s3_meta_request_init_base(
 
     if (meta_request->fio_opts.should_stream && meta_request->fio_opts.disk_throughput_gbps == 0) {
         /* If disk throughput is not set, set it to the default. */
-        meta_request->fio_opts.disk_throughput_gbps = g_default_throughput_target_gbps;
+        meta_request->fio_opts.disk_throughput_gbps = client->throughput_target_gbps;
     }
 
     /* Set up reference count. */

source/s3_util.c

@@ -64,10 +64,31 @@ const struct aws_byte_cursor g_user_agent_header_unknown = AWS_BYTE_CUR_INIT_FRO
 
 const uint32_t g_s3_max_num_upload_parts = 10000;
 const size_t g_s3_min_upload_part_size = MB_TO_BYTES(5);
-const size_t g_streaming_buffer_size = MB_TO_BYTES(8);
+
 const double g_default_throughput_target_gbps = 10.0;
-/* TODO: disable this threshold until we have a better option for threshold */
-const uint64_t g_streaming_object_size_threshold = UINT64_MAX;
+
+/**
+ * Streaming buffer size selection based on experimental results on EBS:
+ *
+ * - Too small buffer sizes (e.g., 16KiB) impact disk read performance,
+ *   achieving only 6.73 Gbps throughput from EBS.
+ * - Too large buffer sizes cause network connections to starve more easily
+ *   when disk reads cannot provide data fast enough.
+ * - 1MiB buffer size provides optimal balance: sufficient disk read throughput
+ *   while maintaining reasonable retry rates due to connection starvation.
+ */
+const size_t g_streaming_buffer_size = MB_TO_BYTES(1);
+
+/**
+ * The streaming approach reduces memory consumption without introducing unexpected errors
+ * or performance degradation.
+ *
+ * We start streaming for objects larger than 1TiB, with plans to lower this threshold in future iterations.
+ *
+ * The 1TiB threshold was chosen to minimize the blast radius of this behavioral change
+ * while still providing meaningful memory usage improvements for large objects.
+ */
+const uint64_t g_streaming_object_size_threshold = TB_TO_BYTES(1);
 
 /**
  * TODO: update this default part size 17/16 MiB based on S3 best practice.