MCPcopy Create free account
hub / github.com/apache/arrow / MakeFromNetworkMetrics

Method MakeFromNetworkMetrics

cpp/src/arrow/io/caching.cc:49–131  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

47}
48
49CacheOptions CacheOptions::MakeFromNetworkMetrics(int64_t time_to_first_byte_millis,
50 int64_t transfer_bandwidth_mib_per_sec,
51 double ideal_bandwidth_utilization_frac,
52 int64_t max_ideal_request_size_mib) {
53 //
54 // The I/O coalescing algorithm uses two parameters:
55 // 1. hole_size_limit (a.k.a max_io_gap): Max I/O gap/hole size in bytes
56 // 2. range_size_limit (a.k.a ideal_request_size): Ideal I/O Request size in bytes
57 //
58 // These parameters can be derived from network metrics (e.g. S3) as described below:
59 //
60 // In an S3 compatible storage, there are two main metrics:
61 // 1. Seek-time or Time-To-First-Byte (TTFB) in seconds: call setup latency of a new
62 // S3 request
63 // 2. Transfer Bandwidth (BW) for data in bytes/sec
64 //
65 // 1. Computing hole_size_limit:
66 //
67 // hole_size_limit = TTFB * BW
68 //
69 // This is also called Bandwidth-Delay-Product (BDP).
70 // Two byte ranges that have a gap can still be mapped to the same read
71 // if the gap is less than the bandwidth-delay product [TTFB * TransferBandwidth],
72 // i.e. if the Time-To-First-Byte (or call setup latency of a new S3 request) is
73 // expected to be greater than just reading and discarding the extra bytes on an
74 // existing HTTP request.
75 //
76 // 2. Computing range_size_limit:
77 //
78 // We want to have high bandwidth utilization per S3 connections,
79 // i.e. transfer large amounts of data to amortize the seek overhead.
80 // But, we also want to leverage parallelism by slicing very large IO chunks.
81 // We define two more config parameters with suggested default values to control
82 // the slice size and seek to balance the two effects with the goal of maximizing
83 // net data load performance.
84 //
85 // BW_util_frac (ideal bandwidth utilization): Transfer bandwidth utilization fraction
86 // (per connection) to maximize the net data load. 90% is a good default number for
87 // an effective transfer bandwidth.
88 //
89 // MAX_IDEAL_REQUEST_SIZE: The maximum single data request size (in MiB) to maximize
90 // the net data load. 64 MiB is a good default number for the ideal request size.
91 //
92 // The amount of data that needs to be transferred in a single S3 get_object
93 // request to achieve effective bandwidth eff_BW = BW_util_frac * BW is as follows:
94 // eff_BW = range_size_limit / (TTFB + range_size_limit / BW)
95 //
96 // Substituting TTFB = hole_size_limit / BW and eff_BW = BW_util_frac * BW, we get the
97 // following result:
98 // range_size_limit = hole_size_limit * BW_util_frac / (1 - BW_util_frac)
99 //
100 // Applying the MAX_IDEAL_REQUEST_SIZE, we get the following:
101 // range_size_limit = min(MAX_IDEAL_REQUEST_SIZE,
102 // hole_size_limit * BW_util_frac / (1 - BW_util_frac))
103 //
104 DCHECK_GT(time_to_first_byte_millis, 0) << "TTFB must be > 0";
105 DCHECK_GT(transfer_bandwidth_mib_per_sec, 0) << "Transfer bandwidth must be > 0";
106 DCHECK_GT(ideal_bandwidth_utilization_frac, 0)

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected