[swift-server-dev] [swift-dev] Performance Refinement of Data
Philippe Hausler
phausler at apple.com
Tue Nov 29 09:57:18 CST 2016
From the beginning of the re-design part of the thoughts were specifically targeted at tuning for server, desktop and mobile environments.
To touch on a few of the specifics that I was focusing on:
Parsing data should be fast - this means a couple of things should be really quick to access. Namely of which accessing subranges, both single and length based ranges should boil down to a memmove or a load from an offset respectively. Under the hood NSData uses memory move/copy operations to access bytes in the fastest manner; or allow the developer to access the internal bytes pointer directly. The latter is of course dangerous and we have much safer mechanisms in swift to deal with that type of access (via closure etc).
Furthermore another very common task is to build up buffers of data (appending small chunks; buckets in the brigade in the parlance of Apache). There is a caveat here of course; if the end result will not be needed to be contiguous DispatchData may be a better type in the end for this but that of course depends on your usage. There is a reason why we have both contiguous data (NSData) and discontiguous data (_NSDispatchData/OS_dispatch_data/DispatchData). So those portions aside of differing usage patterns I will focus on the contiguous case since that is what my proposal is about. The reallocation phase for NSData and Data are almost identical algorithms that account for failure patterns and also aim to avoid memory fragmentation (a big problem for servers that may not be running with a compacting malloc). One of the key advantages for Data versus [UInt8] is that Array must account for varying strides and potentials of retains/releases whereas Data does not need to do this. I did not attach the benchmarks since they were comparing totally different storage mechanisms but Data is either on par or faster than [UInt8] since it is specialized for unsafe pointer access and never has to store objects or other strides - it is just a safe wrapper for UnsafeMutableRawPointer effectively.
In my iterations of this latest version of Data I used a technique where I had a target assembly execution pattern that I tweaked the code until it came as close as possible to that pattern. This development loop (paired with real world scenarios in my profiling) allowed me to suss out cases where we had a number of cases of extra retains of the backing store (which cause atomic operations) or unneeded copies that occurred when the CoW semantics got confused because of complicated code and abstractions. The new version is not only more maintainable (modulo some repetition that was caused because of attempting to avoid closures unless absolutely needed) but also more grokable IMHO (mileage may vary).
The CoW overhead in the new Data implementation only copies when it absolutely needs to. This means that if in a server scenario it has a payload of Data that is being handed around it never needs to copy the bytes of the data and most of the wrapper structure can be stack allocated. Which I think meets the goal of avoiding copies.
Range slicing is now considerably more efficient (I can build some more graphs illustrating that).
And per the last commentary of byte per byte access: please refer to the subscripting graphs - the new version is about 2x faster on my victim machine (old: ~30ns/call new: ~15ns/call)
Let me know if there are other specific examples of usage for bytes access that I should build up graphs for that will illustrate the performance improvements. Overall I think it is a incremental improvement of performance without needing to change APIs which might be able to be applied to some other key classes to get better performance for iOS/macOS and Linux on both devices as well as servers.
> On Nov 29, 2016, at 6:12 AM, Helge Heß via swift-server-dev <swift-server-dev at swift.org> wrote:
>
> On 28 Nov 2016, at 20:04, Alex Blewitt via swift-server-dev <swift-server-dev at swift.org> wrote:
>> This might be of interest to the swift-server-dev community who aren't on either of the other lists.
>
> Thanks.
>
> Hm, I don’t know. Something I don’t quite get is why this is even required. What is really different between that `Data` and a `[ UInt8 ]`? Just the bridging to NSData (which we don’t have on Linux), and couldn’t all that bridging stuff described be done against a special [UInt8] array-backend class too?
>
> The performance patterns attached are focused around a single data object. Also: "Data does not follow this pattern. Often it is sourced from files on disk (which could be exceedingly large)”.
> For the server I think the focus needs to be quite different.
>
> So what are the requirements for a ‘byte bucket’ data structure on the server. Some stuff I can think of:
>
> - There are a lot (“millions") of read and write calls with
> tons of small byte buffers (essentially socket buffer
> sizes, but when framing has been processed even smaller ones ...)
>
> - Copying of those buffers should be avoided as much as possible
> - Ideally it should be possible to reuse such buffers (instead
> of malloc/free’ing them at a high rate)
>
> - It is rather rare that the buffers in servers are itself
> modified, usually you create a new, transformed buffer (or
> other object, like a String or Int) out of existing ones
> - except for one operation: you often slice them (e.g. when
> encountering a HTTP chunk marker, or some other framing in
> a protocol or data format).
> - that might imply that the CoW overhead - even if small -
> may be unnecessarily high
> - instead there needs to be a focus an dealing with ranges
>
> - It should be highly efficient to create String’s from such
> byte buckets (arrays). while it should be delayed to create
> such to the latest point possible, it quite often will be
> the API the top-level framework user is expecting (JSON,
> query-strings, XML, whatever).
>
> - There needs to be something to represent arrays of those
> buckets, Apache calls them brigades. One could use Arrays
> for such, but maybe they should be lists instead.
> (this is required to avoid copying split buffers into
> larger ones, and is supported by writev/readv)
>
> - Except for some special cases like caches - unlike mentioned
> in the proposal - server apps should rarely ‘source byte
> buckets from disks’.
> Instead of loading the buckets into memory, maybe there
> should be a way to have byte-buckets which are files.
> NSData could do mmap’ed files, what you more often really
> want in the server context is `sendfile()` though.
> - some kind of a ‘FileDescriptorBackedBuffer’ protocol
>
> - Scanning of buffers should be really fast - and by that I
> mean C `index()` like fast, not `for byte in bytes {}` array
> fast. But that is desirable for [UInt8] too :-)
>
> There is probably more, but I think the stuff above are things that should be addressed :-)
>
>
> It may also be worth considering whether that ‘buffer’ data structure should only be able to carry bytes, or other payloads as well (along the lines of http://noze.io/noze4nonnode/#streams). But maybe that is too ambitious for v1 :-)
>
> hh
>
> P.S.: This may be good reading material: http://www.apachetutor.org/dev/brigades
>
>> Alex
>>
>>> Begin forwarded message:
>>>
>>> From: Philippe Hausler via swift-dev <swift-dev at swift.org>
>>> Subject: [swift-dev] Performance Refinement of Data
>>> Date: 19 November 2016 at 19:10:32 GMT
>>> To: swift-dev at swift.org, swift-corelibs-dev at swift.org
>>> Reply-To: Philippe Hausler <phausler at apple.com>
>>>
>>> Performance Refinement of Data
>>> Introduction
>>> In Swift 3 the Foundation team introduced a new structural type to represent NSData exposed into Swift.
>>>
>>> Data allows developers to interact with binary data with value semantics, which is often more appropriate than using pointers like UnsafeMutablePointer. Having an encapsulating type to abstract the common operations is often quite advantageous to tasks like parsing, where types like String may not be appropriate for or have hidden performance "gotchas".
>>>
>>> Data can easily be a critical point of performance for many tasks. Data is an appropriate common currency of transacting a safe managed buffer of bytes that interoperates well with existing Cocoa APIs. This means that it should be tuned for performance as much as possible.
>>>
>>> Motivation
>>> There are several outstanding performance improvements which can be made to Data; issues regarding Data's performance have been raised with the Foundation team both publicly and privately, and we would like to address those.
>>>
>>> Data should be as fast as possible. Currently, most of the backing of Data is implemented in Foundation, while being quite fast for reallocations and other operations, this means that calls are made between Swift and Objective-C even for simple things like count for every access.
>>>
>>> This Swift–Objective-C boundary means that no inlining can be performed across it; even when we have full control over the backing implementation and the caller, what would normally be just a single offset load instructions ends up becoming many just for the benefit of an objc_msgSend (not to mention ARC operations). Even though the calls are heavily optimized, they will never be as fast as a single instruction.
>>>
>>> Proposed solution
>>> In order to make Data as fast as possible the implementation needs to be inlined; since that is one of the best performance optimizations that Swift can offer. To do this it requires a re-think of how Data and NSData interact. This means that the structure Data will need to adopt certain attributes that will allow it to be inlined into the call-sites. The function dispatch overhead will be reduced but also optimization passes like cold paths and branch elimination can be possibilities for the compiler to do the best possible thing for the call site.
>>>
>>> Data will adopt the annotation @inline(__always) in key locations and use a non-Objective-C backing class to store the pointer and length (and other internal ivars). That backing object will allow for a reduction of capture points as well as avoid extra retain/releases that were added for mutation detection.
>>>
>>> Instead of using _MutablePairBoxing (which uses closures to map invocations to references or apply mutations with copy on write semantics) the new backing implementation can easily be applied with copy on write semantics without any Unmanaged "dancing". That "dancing" complicates code and it can be made considerably simpler. Furthermore, avoiding this "dance" can reduce the calls to retain and release down to zero in the application of mutations in unique referenced cases as well as mapping non mutating invocations to backing storage.
>>>
>>> Subclassing the reference type NSData is still something that Foundation should support for the wrapping of the reference type in a structure. This effectively means there are five types of backing for Data: Swift-implemented, immutable NSData, mutable NSMutableData, custom subclasses of NSData, and custom subclasses of NSMutableData. These specific cases are delineated to offer the most efficient inline cases possible.
>>>
>>> Since Foundation can enforce a no dynamic dispatch needed contract with itself in the cases of the standard class cluster members of NSData and NSMutableData Foundation can assure these cases are acceptable to not need dynamic dispatch for every time bytes or length are accessed and the values can be cached until the data is mutated or disposed of. In the cases where a subclass is used of course all bets are off and every point requires dynamically calling out.
>>>
>>> In short this will mean that fetching the count of a Data can be optimized to a single branch and load from an offset and this same optimization can be applied to many other methods on Data.
>>>
>>> Bridging to and from Objective-C
>>> Many of the sources that Data is derived from are sourced from the SDKs written in Objective-C. For other types like Array,Set, Dictionary, or String the objects returned are many times not very large. Arrays may have a handful of objects, strings may only be a few hundred characters and so on. In these small cases it makes sense to "eagerly" bridge those reference types into a more inline-able version (there are exclusions to this but in general it is most often the case).
>>>
>>> Data does not follow this pattern. Often it is sourced from files on disk (which could be exceedingly large) or results from network transactions of downloads. These cases would definitely suffer from having an "eager" O(n) bridge; due to not only memory allocation duplications to hold both backing stores but also to the byte copy from the reference type to the value type. Data should be fast no matter where it came from unless it is truly unknown on it's dynamic dispatch requirements.
>>>
>>> To build a Data that is fast for inline optimizations the bytes pointer and length need to be cached for the duration of the object. When as is used to bridge a custom reference to Data dynamic dispatch must occur on every call to count and every time bytes are touched but if the Data is known to be obtained from a source that we can control the dynamic dispatch expectations that dynamic dispatch can be elided and behavior can be preserved by mimicking the Objective-C implementation in Swift.
>>>
>>> Bridging in the other direction also has some distinct performance optimizations that can be taken advantage of as well.
>>>
>>> When the lifespan of the callout to Objective-C is well known the cases of Swift constructed Data can easily pass a NSData with a no-copy of the backing buffer. It is the responsibility of the Objective-C APIs to appropriately either not directly retain past the scope of the call or copy in the cases of long lasting references. Any Objective-C method or function that takes a NSData and just retains or unsafely stores it past the function callout is likely incorrect and has bugs no matter the language it was invoked in. This case where the Data is created in Swift to bridge it only needs to allocate the wrapper NSData but no O(n) copy needs to occur (unless it is holding a reference as previously stated).
>>>
>>> The final case of bridging is when a Data is obtained from Objective-C and then passed directly back to Objective-C. The compiler has potentials of optimizations in direct callout cases such as returnsAData() as NSData with "peephole" optimizations but these are only really enforceable in limited scope (sometimes just the same line of code). Since the backing store can hold a reference to the reference type the bridge method (when not mutated) in those cases can pass that back over to Objective-C. For mutated versions a copy of that mutated version can be passed along as well (taking advantage of any optimizations the dynamic dispatch affords for calls to copy).
>>>
>>> Detailed performance breakdown
>>> Each graph below is a comparison between the Swift 3 Data and the new version of Data for each of the inline cases. The horizontal axis in each graph represent N and the vertical axis in each graph represents the sampled duration in nanoseconds. Each data set in the plots are an average over 100 (unless otherwise specified) per value of N. The attached graphs were generated from optimized builds on a Mac Pro (Late 2013) 3.5 GHz 6-Core Intel Xeon E5 with 16 GB 1866 MHz DDR3.
>>>
>>> func createSampleData(ofLength N: Int) -> Data {
>>> var buffer = [UInt8](repeating: 0, count: N)
>>> return buffer.withUnsafeMutableBytes { (buffer: UnsafeMutableRawBufferPointer) -> Data in
>>> arc4random_buf(buffer.baseAddress!, N)
>>> return Data(bytes: buffer.baseAddress!, count: N)
>>> }
>>> }
>>>
>>> func createSampleDataReference(ofLength N: Int) -> NSData {
>>> var buffer = [UInt8](repeating: 0, count: N)
>>> return buffer.withUnsafeMutableBytes { (buffer: UnsafeMutableRawBufferPointer) -> NSData in
>>> arc4random_buf(buffer.baseAddress!, N)
>>> return NSData(bytes: buffer.baseAddress, length: N)
>>> }
>>> }
>>>
>>> func createSampleArray(ofLength N: Int) -> [UInt8] {
>>> var buffer = [UInt8](repeating: 0, count: N)
>>> buffer.withUnsafeMutableBytes { (buffer: UnsafeMutableRawBufferPointer) -> Void in
>>> arc4random_buf(buffer.baseAddress!, N)
>>> }
>>> return buffer
>>> }
>>>
>>> Accessing count
>>>
>>> This should be a O(1) operation. The y axis is measured in nanoseconds sampled over 100000 iterations.
>>>
>>> // setup
>>> let data = createSampleData(ofLength: N)
>>> // start measuring
>>> _ = data.count
>>> // end measuring
>>>
>>> <access_count.png>
>>>
>>> Subscripting
>>>
>>> This should be a O(1) operation. The y axis is measured in nanoseconds sampled over 100000 iterations.
>>>
>>> // setup
>>> let data = createSampleData(ofLength: N)
>>> // start measuring
>>> _ = data[index]
>>> // end measuring
>>>
>>> <getting_subscript.png>
>>>
>>> // setup
>>> var data = createSampleData(ofLength: N)
>>> // start measuring
>>> data[index] = 0x00
>>> // end measuring
>>>
>>> <setting_subscript.png>
>>>
>>> Appending
>>>
>>> This should be a O(N) operation
>>>
>>> // setup
>>> let dataToAppend = createSampleData(ofLength: N)
>>> var data = Data()
>>> // start measuring
>>> data.append(dataToAppend)
>>> // end measuring
>>>
>>> <append_n_bytes_with_data.png>
>>>
>>> // setup
>>> let arrayToAppend = createSampleArray(ofLength: N)
>>> var data = Data()
>>> // start measuring
>>> data.append(contentsOf: arrayToAppend)
>>> // end measuring
>>>
>>> <append_array_of_bytes.png>
>>>
>>> The new version is still O(N) just a much smaller constant multiplier.
>>>
>>> var data = Data()
>>> // start measuring
>>> for _ in 0..<N {
>>> data.append(contentsOf: [0xFF, 0xFE, 0xFD, 0xFC, 0xFB, 0xFA])
>>> }
>>> //end measuring
>>>
>>> <append_n_arrays.png>
>>>
>>> Replacing sub ranges
>>>
>>> // setup
>>> var data = createSampleData(ofLength: N)
>>> // start measuring
>>> data.replaceSubrange(0..<N, with: replacement)
>>> // end measuring
>>>
>>> <replace_entire_subrange.png>
>>>
>>> // setup
>>> var data = createSampleData(ofLength: N)
>>> // start measuring
>>> data.replaceSubrange(0..<min(N, 5), with: replacement)
>>> // end measuring
>>>
>>> <replace_fixed_subrange.png>
>>>
>>> Growth of count
>>>
>>> // setup
>>> var data = Data()
>>> // start measuring
>>> data.count = N
>>> // end measuring
>>>
>>> <grow_small.png>
>>>
>>> // setup
>>> var data = Data()
>>> // start measuring
>>> data.count = N
>>> // end measuring
>>>
>>> <grow_large.png>
>>>
>>> // setup
>>> var data = Data()
>>> data.count = starting
>>> // start measuring
>>> data.count = N
>>> // end measuring
>>>
>>> <grow_from_mid_to_large.png>
>>>
>>> Bridging to reference types
>>>
>>> This should be a O(1) operation. In bridging to a reference case the previous implementation was a bit faster. The only real extra overhead here is an allocation of the NSData object since the Swift backed Data has no existing reference type to pass along. There are a few extra optimizations that can be done in this path to reduce it by the approximately 150 nanosecond difference. In practice the cases where Data is being bridged back out to Objective-C are usually cases like writing to a file or socket which dwarf that 150 nanosecond differential.
>>>
>>> // setup
>>> let data = createSampleData(ofLength: N)
>>> // start measuring
>>> _ = data as NSData
>>> // end measuring
>>>
>>> <bridge_to_objectivec.png>
>>>
>>> Bridging from reference types
>>>
>>> This should be a O(1) operation
>>>
>>> // setup
>>> let data = createSampleDataReference(ofLength: N)
>>> // start measuring
>>> _ = data as Data
>>> // end measuring
>>>
>>> <bridge_from_objectivec.png>
>>> _______________________________________________
>>> swift-dev mailing list
>>> swift-dev at swift.org
>>> https://lists.swift.org/mailman/listinfo/swift-dev
>>
>> _______________________________________________
>> swift-server-dev mailing list
>> swift-server-dev at swift.org
>> https://lists.swift.org/mailman/listinfo/swift-server-dev
>
> _______________________________________________
> swift-server-dev mailing list
> swift-server-dev at swift.org
> https://lists.swift.org/mailman/listinfo/swift-server-dev
More information about the swift-server-dev
mailing list