Patcher Part 8 : Summary

In this series I described how to build a patcher that can make very good (near minimal size) "coarse grain" patches, and run near maximum possible speed (IO limited).

Posts in the series :

cbloom rants- Patcher Part 1 - Introduction with context and defining terminology and goals
cbloom rants- Patcher Part 2 - Some Rolling Hashes
cbloom rants- Patcher Part 3 - How rsync works
cbloom rants- Patcher Part 4 - Content-Defined Chunking
cbloom rants- Patcher Part 5 - Aside for some proofs
cbloom rants- Patcher Part 6 - Making a patcher from CDC
cbloom rants- Patcher Part 7 - Patcher File IO and Parallelism

To wrap up, some real performance and patch sizes :

Generating patch sizes for three different Fortnite releases on my ThreadRipper :

patcher run time : 20.533 s total_io_bytes=99.326 gB speed=4.837 gB/s
patcher run time : 23.461 s total_io_bytes=100.165 gB speed=4.269 gB/s
patcher run time : 18.366 s total_io_bytes=77.022 gB speed=4.194 gB/s

These times show the total time and net speed (net speed = total bytes over total time), eg. including startup allocs and shutdown frees. These are times just to generate patch sizes ("patcher s" mode), not including writing the patch output files. The three times are for three different data shapes; the first is lots of loose files, patching one file to one previous file, the second is lots of loose files but patching each new file from all previous files, and the first is patching big aggregate tar against previous. (those three styles are interesting because they multi-thread differently).

For sanity check, I verified patch size against rdiff :

On a single 38 GB Fortnite package file :

rdiff with block size set to 1024 :

rdiff_delta fn2401_pakchunk10-XboxOneGDKClient.ucas fn2410_pakchunk10-XboxOneGDKClient.ucas

fn2401_2410_xb1_rdiff_delta_bs1024     927,988,455

vs my patcher with default block size (1024) :

patch size : 903386162 / 38553104896 bytes = 2.34%

Our patch is 2.6% smaller than rdiff. Our patch size should (almost) always beat rdiff by a little because of match extension. ("almost" because there is some randomness and luck in the hash-based CDC splits and chunk matches, so you can get unlucky).

Another rdiff comparison with timing :

On my Intel laptop ; Core i7-8750H CPU @ 2.20GHz , 6 cores (12 hyper)

On a 3.5 GB PDB , block sizes 1024 :

rdiff : fnpdb.rdiff_bs1024     807,066,347
rdiff timerun: 127.729 seconds

patcher : patch size : 792333063 / 3498102784 bytes = 22.65%
patcher run time : 2.681 s total_io_bytes=6.996 gB speed=2.609 gB/s

(patcher "s" mode, only generating patch size not writing output)

Again we're similar size but slightly smaller, as expected. Rdiff takes 127.7 seconds to our 2.7 seconds, so yes good patches can be made much faster than current tools. To be fair, many of the techniques we use could also be applied to speed up rdiff; the rdiff/rsync algorithm is not inherently a terrible way to generate patches and could be ~10X faster than it is now. Also rdiff here is writing the signature file to disk, and actually writing the patch file, while my run is writing no output to disk, so it's not apples-to-apples, and of course we are multi-threaded but rdiff is not. So it's by no means intended as a direct comparison of the maximum theoretical speed of the rsync algorithm vs the cdc algorithm, which should be much closer. The point is sort of that all those practical things to make a fast patcher are important and surmountable.

For the record a full profile run of the above 2.7s PDB patcher run :

On my Intel laptop ; Core i7-8750H CPU @ 2.20GHz , 6 cores (12 hyper)

m:\test_data\patcher>patcher s fnpdb\FortniteClient.pdb fnpdb\FortniteClient.pdb.2 -p
patcher built Oct 10 2023, 09:41:23
args: patcher s fnpdb\FortniteClient.pdb fnpdb\FortniteClient.pdb.2 -p
got option : do_profile
detected disk type = ssd
cores_hyper = 12 physical = 6 large_pages = true
io_limit_count=2 cpu_limit_count=12
make patch from fnpdb\FortniteClient.pdb to fnpdb\FortniteClient.pdb.2
FortniteClient.pdb.2: read_and_make_sourcefile: 3498102784 start...
FortniteClient.pdb.2: read_and_make_sourcefile: 3498151936 start...
patch size : 792333063 / 3498102784 bytes = 22.65%
 patch bytes matched :  70053394 / 3498102784 bytes =  2.00%
 patch bytes nomatch : 792232278 / 3498102784 bytes = 22.65%
 patch bytes zeros   : 2635817112 / 3498102784 bytes = 75.35%
patcher run time : 2.681 s total_io_bytes=6.996 gB speed=2.609 gB/s
SimpleProf                       :seconds    calls     count :   clk/call  clk/count
patcher                          : 2.6714        1         1 :   10661.5m  10661.54m
 makepatch_one_file              : 2.6314        1         1 :   10502.1m  10502.10m
  read_and_make_sourcefile_twice : 2.3262        1         1 : 9283832.7k   9283.83m
   read_and_make_sourcefile      : 4.5463        2         2 : 9072217.6k   9072.22m
    make_fragments               : 2.3581      677  6996987k : 13901373.8       1.35
     make_fragments_sub          : 1.7625     5068  4351168k :  1387917.2       1.62
    ComputeWholeFileHash         : 0.3554        2         2 :  709214.9k 709214.94k
  makepatch_one_file_sub         : 0.1575        1  3498102k :  628657.0k       0.18
   make_base_fragment_hash       : 0.1192        1         1 :  475762.4k 475762.35k
  write_one_file_patch           : 0.1444        1 792333063 :  576399.6k       0.73

The bulk of the time is in the "read and make signature" CDC phase which does the boundary scan and hashes each fragment. This can be done on the two input files at the same time so the net CPU time for it is 4.5463s (read_and_make_sourcefile), but the wall time is 2.3262s (read_and_make_sourcefile_twice).

Other applications :

We've looked at patching from just one "previous" file (or set of previous files) to one "new" file, but there are other applications of these techniques. You can do partial transmission like rsync, you can do file system dedupe like ZFS.

I use "patcher s" size mode just as a fuzzy binary diff. If the patch size needed to get from file A to file B is small, it tells me they are very similar.

The way we use "patcher" at Epic is currently primarily for analysis. On most of our important platforms, the actual patches are made by the platform distribution tools, which we cannot control, so we can't actually generate our own patches. We use "patcher" to see what the minimum possible patch size would be with a perfect patch tool, so we can see how much of the patch size is from actually changed content vs. bad platform patch tools. We can also use "patcher" to output a CSV of change locations so that we can identify exactly what content actually changed.

Another application I've used is grouping similar files together for compression (which I call "orderer"). The CDC "signature" provides a quick way to identify files that have chunks of repeated bytes. When making a multi-file solid archive, you gain a lot by putting those files close to each other in the archive. To do that you just take the N files and compute signatures, then try to put them in a linear order to maximize the number of chunk hashes that are equal in adjacent files. Note the rsync/rdiff method works okay (similar big-O speed to CDC method) when doing file-to-file patching, but the CDC method has a big advantage here of being totally symmetric, the expensive work is per-file not per-compare, so when doing a group of N files where they can all match each other, it is a huge win.

You could also use these techniques to find long-distance matches within a huge file which can be used to make sure you find LZ matches (for something like Kraken or ZStd) at very long range (which can otherwise get lost when using something like a cache table matcher, since very old/far matches will get replaced by closer ones). (but realistically if you are doing super huge file compression you should have something specialized, which maybe finds these big repeats first, and then compresses the non-repeating portions in independent pieces for multi-threading) (I believe that "zpaq" does something like this, but haven't looked into it).


Patcher Part 7 : Patcher File IO and Parallelism

In the real world, a lot of the issues for making a very fast patcher are in the practical matters of parallelism and file IO, so let's dig into those a bit.

I believe that it is bad practice to take unnecessarily slow algorithms and just throw them onto threads to make your program fast by using tons of threads. So first we tried to make the patcher algorithms as fast as possible on a single thread, and we got the core operation (CDC "signature" computation) down to 1.5 cycles/byte, but that's still not fast enough to keep up with IO, so we will need parallelism.

The "speed of light" for patcher, the fastest it can possibly ever go, is the time to just do the IO to read the previous & new file sets, and to write the patch output. We want to hit that speed, which means we want to be totally IO bound. A typical current SSD can do around 6 GB/s ; on a 3 GHz CPU that's 0.5 cycles/byte for IO. So naively that tells us we need 3 computation threads running 1.5 cycle/byte work to keep up with IO. (modern PCIe 5 drives can go even faster and would need more computation threads to saturate the IO).

When doing parallelism work, it's useful to think about what is the single-threaded critical path that cannot be parallelized and will limit your speed (even if you had infinite thread count). In this case it's easy, it's just the IO. So as long as we are always doing IO, keeping the disk running at maximum speed, and overlapping CPU work alongside that, we will achieve maximum speed.

The primary operation of the patcher is the computation of the CDC signature, which has the basic form :

read whole file into buffer

scan over buffer doing hash computations, making fragments

Parallelizing over patch sets that consist of lots of small files is trivial, but to parallelize (and crucially, overlap the IO and computation time) over single large files requires interleaving the IO and computation work on individual files. The real world data sets we work on tend to be either single very large files (such as when patching a whole distribution that's packed together with something like tar), or a bunch of files of various sizes, we want to handle all those cases well.

Since IO speed is crucial here, I did some experiments on a couple different disk types on a couple different machines, and I will briefly summarize what I found. Caveats: this is very Windows specific; I use the Win32 OVERLAPPED API. I do not have a modern PCI5 super-fast SSD or a Zen 4 CPU to test on; my fastest SSD is around 6 GB/s, some results may differ on new PCI5 SSD's. I did test on 3 machines : an Intel CPU, a Ryzen Zen 3, and a ThreadRipper, all with both an M2 SSD and a spinning platter HDD. I did not test with SetFileValidData to get true async writes, as that is not practical to use in the real world so is moot.

Summarizing what I found :

  • You can use multiple IO threads to read from SSD's at full speed, but multiple threads reading from HDD are a huge huge disaster. You must use only 1 thread for IO on HDD.

  • Always do reads unbuffered. Using a buffered read causes extra mem copies through the IO buffers, which is a significant speed penalty on fast SSD's. (buffered reads don't hurt on HDD, but it's simpler to just say use unbuffered reads all the time).

  • Use unbuffered writes on SSD. Use buffered writes on HDD. On some of the HDD's on some of my systems, buffered writes were significantly faster than unbuffered (120 MB/s vs 80 MB/s). (to be clear I mean buffered at the Win32 CreateFile HANDLE level, you should never use stdio or your own extra buffering system for fast IO).

  • Use 4-16 MB io chunk sizes. This is small enough to be incremental at reasonable granularity and big enough to run at full speed.

  • For incremental reading of a file on a single thread, do OVERLAPPED async IO and keep two OVERLAPPED structs running at all times, like a double-buffer. That is, fire off two async reads, when the first completes fire it for the next chunk, etc. This ensures you always have a pending async read enqueued to the device, you aren't doing an IO, then going back to your io thread to enqueue the next, leaving the device idle for a while until you get the next chunk requested.

  • SSD's can run reads and writes at the same time at full speed. For example, to do a file copy on an SSD you should run the reads and writes at the same time, which can be done on a single thread use triple-buffering of async/overlapped IO (so you always have an async read and write in progress).

  • Some IO operations (eg. dir listing) benefit from being heavily multi-threaded (running on all cores, not just 1), because they are mostly CPU bound (working on data structures that are often in memory already). For the real bandwidth heavy work (reading,writing), lots of threads doesn't help. You can get full speed with only 1 IO thread for HDD, and 2 for SSD.
On the 3 machines and 6 disks that I tested on, this recipe gave near-optimal speed in all cases. That is by no means a thorough survey, and different disks or OS or config situations may give different results. My goal was to find a simple standard recipe for fast IO that doesn't involve a lot of per-machine tweaking, which could easily get over-trained for my specific machines. I also believe in using as few threads as possible that get you to full speed.

Out in the wild you can have funny issues affecting IO perf, such as running in some kind of VM or container, running with a virtual file system driver from a virus scanner, files on network drives, etc. Timing IO can be tricky because of the effects of OS buffers, and writes returning from your API call before they actually go to disk. Some disks are fast at first them go into a slower mode when used heavily, either due to caches or thermal throttle.

Detecting SSD vs HDD is pretty simple on Windows; it's in cblib ("DetectDriveType") as is a basic double-buffered OVERLAPPED reader and triple-buffered copier ("win32_sync_copy").

The basic threading model patcher uses is this :

Run 1 file at a time on HDD, 2 at a time on SSD

On each file, do async OVERLAPPED IO in chunks for incremental IO

As each chunk is done, kick off CPU work to process that chunk (compute "signature")

The "signature" finds CDC boundaries and computes hashes on each fragment. We do this in 16 MB chunks, which means we get artificial cut points (not content-determined) at the 16 MB IO chunk boundaries. You could just ignore that, as it's a small fraction of the total size, but instead what I do is after an adjacent pairs of chunks is done, I delete the fragments that were made near the boundary (two on each side of the boundary) and re-find CDC boundaries in that small region.

On an SSD at 6 GB/s and CPU at 3 GHz, the rough times are 0.5 cycles/byte for IO and 1.5 cycles/byte for signature building. So the timeline lookes like :

different 16 MB chunks labelled A,B,C
time is on the X axis

work: BBBEEE
work:  CCCFFF

That is, three time units of work on CPU worker threads per IO chunk, so we need three threads doing computation to keep up with the IO speed.

The signature computation for the "previous" and "new" file are the vast majority of the CPU work, but once that is done we have to build the hash table and then match against that hash table, which is pure CPU work. During this phase, we can be running the next file in the set, doing its IO phase.

To do that easily, I use a simple semaphore to throttle the IO threads, rather than a true dedicated IO thread. (I think a true IO thread is probably a better model, with followup work spawned on IO completion, but it makes the code much less linear-imperative, so the semaphore method while a bit less efficient is much easier to read and maintain). The IO semaphore means only 1 thread can be running IO at a time (as required for HDD, or 2-3 threads for SSD), but which thread that is changes.

So what we actually do is :

parallel for on files , (memory limited and io_sem limited, see later)

take io_semaphore { read file1 and make signature, incrementally, uses 3 worker threads }
take io_semaphore { read file2 and make signature, incrementally, uses 3 worker threads }

^ these two can run at the same time if two io_sem decrements are available

now the io_sem is released, so the next file can start its io work immediately

do cpu work :
build hash table
match against hash table
construct patch

release memory, which unblocks the memory limitted queue

For threading primitives, I used the C++ std::async, as well as Microsoft's concrt/ppl. (I did this sort of as a learning experiment to try some not-quite-modern C++ threading code instead of using the mechanisms I had written for Oodle).

On Windows, std::async and concrt/ppl are both built on the Windows thread pool. When you start async tasks or do a parallel_for, they take threads from the thread pool, or create new ones if necessary. Sadly on gcc/Linux, std::async starts a new thread for each async task (no thread pool), which is no good, and means we can't use std::async to write cross platform code.

The Windows thread pool mostly works okay for our purposes. Thread pools solve the "wait from worker" problem by switching to another thread in the pool when you wait on a worker, which keeps a task running on all cores at all times. (as opposed to coroutine yields, or fibers, or "pop on wait", which are alternative solutions to "wait from worker"). This is mostly okay but requires some care. When you do a wait from a pool thread (such as waiting on an IO to finish, or waiting on a mutex/critsec), it can cause a new thread to start up, so that something runs while you stall. Then when your wait is done, your thread can start running again, but the new thread that was started may still exist. This can cause the pool to get many more threads than cores, and give you extreme over-subscription.

As an example of a terrible way to use a thread pool, consider the common pattern of doing an IO to read a whole file, then doing some processing on that file :

parallel_for over all files :
  phase1: read whole file into buffer (wait on io completion)
  phase2: act on buffer, doing some computation

Say you have something like 32 cores and 1000 files to process. The parallel_for will make 1000 tasks and send them to the thread pool to execute. Initially the pool will kick off a task to a worker thread for each core (32 of them). Each of those tasks will try to start a file IO then wait on it. So those 32 threads will all go to sleep. The thread pool will see it has no threads running on the cores, but lots of pending tasks, so it needs to make more threads; it will make 32 more threads, all of which will block on IO and go to sleep. Eventually we wind up with 1000 threads all sleeping on IO. Later, the IO's start to finish and the tasks are woken up to move onto phase2 for doing some computation on the IO results. The problem is we now have 1000 threads that all want to run and do CPU work.

This is just in the nature of the way a thread pool addresses the "wait from worker" problem. (note that "pop on wait" and "deep yield" also have their own bad patterns and are by no means a magical solution either, it's simply a messy problem that will always have some bad cases). There are some fudges that make it not actually this bad, for example you can set a maximum thread count in the pool to be something like 2X the core count, but the better solution is to just not use the thread pool in that way. In general, it works well if you avoid waiting from tasks, and instead use followup tasks that trigger from task completions (eg. dependencies).

Specifically in the "patcher" case we do have this common pattern of do some IO and then kick off CPU work to act on that IO. So we can use some better primitives for that. For example we can make a "parallel_for" that loads the file contents one by one, or using 2 threads, and then kicks off the followup cpu-only work :

parallel_for over all files using 1 or 2 IO threads :
  phase1: read whole file into buffer (wait on io completion)
  start async task on thread pool of all cores :
    phase2: act on buffer, doing some computation

Another common useful pattern that I use is to have a parallel_for that pre-takes a critsec or semaphore. Say you have some task that you know needs to immediately take a critsec at the start of the task :

parallel_for :
    enter critsec C
      do X
    leave critsec C
    do more work Y

This will have a similar problem to the IO task on a threadpool. You will start too many threads, they will all block on the critsec C, then once they all get past phase1, you will have too many threads running phase2.

One solution is to have a parallel_for that only dispatches tasks with the critsec entered :

single threaded for :
  enter critsec C
  start async task on thread pool :
    // critsec C is already owned by me
      do X
    leave critsec C
    do more work Y

Note that when the async task leaves critsec C, the single threaded for loop can then step to the next item in the list while the async task proceeds with "work Y". So we get the desired result that "work Y" can run on the thread pool over all cores, but we aren't starting threads just to park them in a wait on the critsec. (also note that this has the non-ideal property of going back to the calling thread to activate followup work, which we would rather do with a dependency system to do direct handoff, but that requires more complex mechanisms).

A related issue is that we sometimes don't want to go parallel over all cores, because we are working with large data sets and we can exceed RAM if we go too wide on our parallelism. It's catastrophic for performance to exceed RAM and go to swap file, so we would much rather dial down parallelism. eg. we often work on dirs containing many 4 GB files at Epic; we'd like to do a parallel_for over files on all of those, but only as long as we fit in memory, which on something like a ThreadRipper can be lower than core count.

To do that I use a specialized memory limited parallel for. Each task to be run in the memory_limited_parallel_for must report its memory use before running. At the start of the parallel_for, the free physical memory is queried; tasks will be run such that their total reported memory use is <= the initial free physical mem. The parallel_for then starts tasks, up to a max of core count running at a time, and only if the memory use fits in current available count. I use a simple greedy scheduler, which runs the largest memory use task which can fit in the current available. This is not optimal but works okay in practice.

(in "patcher", memory use and run time of tasks are both proportional to file size, so larger mem use tasks will take longer, so we want to start the largest mem use tasks as early as possible. Also when no tasks are running, we always run the largest mem use task, even if its reported mem use exceeds the total available.)

Something that you find whenever you work with huge amounts of memory is that simply doing the VirtualFree() to free memory is incredibly slow. In the patcher, on a 30s run, fully 5s was in a VirtualFree which was on the critical path.

Some quick notes about VirtualAlloc/VirtualFree time, then I'll describe how I solved it in patcher.

VirtualAlloc takes no time and immediately returns you a pointer in virtual address space, but has not actually yet mapped pages of memory to your process. That doesn't happen until you actually touch those pages. So the allocation time shows up as being very fast and the actual time for it is distributed around your code as you use those pages.

(some people don't like that, so they will start up an async task to scan through the pages and touch them all right after the VirtualAlloc. That may or may not help your net process time; in some cases it's better to let that work happen on first touch (eg. if you don't always actually use all the memory you requested). One big advantage of doing the separate action to touch pages is it's easy to parallelize that work, and it removes that first-page-touch time from profiles of your other functions, which can make optimizing easier.) (also large pages make this all better, but aren't practical to use in Windows because they require group policy tokens).

VirtualFree is blocking and slow; it has to go through all the pages mapped to your process and give them back to the system. First note that if you didn't actually touch any of the pages, then VirtualFree will be fast. It is only slow if the pages have actually been mapped to your process. If you just do a test app that does "VirtualAlloc then VirtualFree" without touching pages, everything will seem fast.

(there are also issues that can arise with the Windows memory-zeroing of pages which we will not get into here)

You might think that just exiting without freeing, and letting Windows clean up the leaks in ExitProcess would save you from the time to VirtualFree, but that is not the case. Windows ExitProcess blocks on freeing all the memory you have allocated, so the net process time is not reduced by leaking the memory. (TerminateProcess is the same as ExitProcess in this regard). You have to measure the time for your process to return to the calling process.

To be very concrete :

int main(int argc,const char *argv[])
    SIZE_T size = 32LL<<30;
    void * mem;


    if ( do_touch )

        char * ptr = (char *)mem;
        char * end = ptr + size;

            *ptr = 0;
            ptr += 4096;

    if ( do_free )


    return 0;

Then we can look at net process time with do_touch and do_free toggled :

c:\src\testproj\x64\release>timerun TestProj.exe
Timer : alloc : 0.000158 s
Timer : touch : 3.098679 s
Timer : free : 2.214244 s
timerun: 5.332 seconds
    -> true allocation time is only seen when you touch pages

c:\src\testproj\x64\release>timerun TestProj.exe
Timer : alloc : 0.000162 s
Timer : touch : 3.089498 s
timerun: 5.433 seconds
    -> 2.4s in process time not in the timers
    ExitProcess *does* stall on the free

c:\src\testproj\x64\release>timerun TestProj.exe
Timer : alloc : 0.000168 s
Timer : free : 0.000082 s
timerun: 0.013 seconds
    -> free is fast if you don't touch pages

So we understand the issue a bit, what can we do about it?

Well, when we free memory, we don't usually need to block on that operation. The next lines of code don't depend on the free being fully done, we're just trying to say "I'm done with this, free it sometime". So the obvious thing to do is to just launch the free off on an async task, which we don't block on. We just kick off the async task and let the free complete whenever it manages to do so. (I call this a "detached" async task, when the handle to it is just dropped and it should delete itself when done).

There is one exception to that, which is the next time we need to allocate a lot of memory, we don't want that to fail (or go to page file) because we had a detached free that was still pending. eg. you're on a 128 GB system, you alloc 100 GB then free it, then go to alloc 100 GB again, you actually now do want that preceding 100 GB free to be done before your next alloc.

This is a problem we can encounter in real runs in patcher, because we are working with very large data sets near RAM size. To address that I use what I call "robust detached frees".

For "robust detached frees" we still kick the free off on an async task, but we don't just forget the task handle, instead we keep a list of pending frees. As long as we never try to do a big alloc again, then those frees just detach and complete whenever they get around to it. But, if we try to do an alloc that would cause the net committed memory to exceed physical memory size, then we see if there are any pending frees that we did before and block on those before doing the alloc.

So further small allocs won't cause us to block on pending frees, but if we do try a big alloc we will wind up blocking. This typically gets us the benefit of not blocking on the frees on our critical path.

Insert smug self-congratulatory conclusion here.


Patcher Part 6 : Making a patcher from CDC

We have a scheme to cut our file into content-defined chunks . So let's use that to make a patcher.

For each file, we can construct a "signature" analogous to the rsync signature (which is a hash of chunks of constant length at regular intervals). Our signature is all the CDC chunk locations, and then a hash of each chunk. We will use the hash to look up the contents of each chunk; it is not a hash we need to roll, and we don't need it to be cyptographic. A good hash to use here is XXH3 .

CDC file signature :

vector of :
    chunk length (as determined by CDC)
    hash of chunk (64,128, or 256 bits, eg. from XXH3)

chunk position = sum of previous lengths

Basically we will look at these signatures and match chunks that have the same hash. There are a few details to cover.

An optional step that I add in my patcher is to find long runs of zeros and give them their own chunk. Long runs of zeros are something that happens in the real world quite a lot (eg. in PDB, EXE, tar, and Unreal's packages), and isn't handled great by the rest of the system. In the worst case, runs of zeros can be a degenerate point in the CDC chunk-finding rolling hash. While I wound up choosing the table-based rolling hash that does not have that degeneracy, it's easier to test hash options if that is not an issue (if you don't special case zero runs, then whether or not a hash has a degeneracy on zeros affect dominates everything else about the hash).

The CDC signature system also doesn't handle it well when runs of zeros change length. Say you have a "before" file that has a zero run of 1000 bytes, and in the "after" file there's a zero run of 1001 bytes. The CDC rolling hash may put a cut point at the beginning of the zero run and one at the end, but the chunk containing the zero run is not the same before/after so doesn't match. By special casing the zero runs, we can send changes in the length of zeros without causing any patch mismatch.

In theory you might want to find other repeating patterns, but in practice zero runs are the important common case (many file formats use zeros to pad elements to some alignment). Note also that the scan for run patterns must be extremely fast or it will slow down the patcher. I use a minimum length zero run of at least 32 bytes, which can be checked for by looking at 16-byte aligned positions for 16 bytes of zeros (any run of zeros of length >= 31 will contain 16 zeros at a 16-byte aligned position).

So my modified CDC scheme is :

Find long zero runs
Make "zero-run" chunks where there are long zero runs

In between zero run chunks, do the CDC rolling hash to find chunk boundaries
Make chunks from those boundaries
Compute the data hash (XXH3) of those chunks

Okay, so we have a "signature" of the previous & new files, let's make the patch.

First we take all the chunks of the "previous" file (or possibly multiple files if you want to patch from a previous install of multiple files), we take the data hash of all non-zero-run chunks and add them to a hash table.

Some care is needed with the hash table to make sure this is fast. Pre-size the hash table for the chunk count so it doesn't need to resize as you add. Use an "open addressing" reprobring hash table where the entries are stored directly in the table, no pointer indirection. Do not use the STL hash_map. Make the entry as small as possible to reduce memory size. Because the hashes we are inserting are already very well scrambled, the hash table should not do any extra work to do additional munges of the hash. Since our "key" is already a hash, don't compute or store a separate hash of the key. It also helps to prefetch ahead during the adds. See cblib for one example of a good hash table, though something specialized to purpose will always be best.

Note the same hash value may occur many times, if your file has chunks of data that repeat. It is optional whether you add multiple occurance of the same chunk contents (which occur at different locations) or not. If you want to do a patcher that requires data locality, then you might want to go ahead and add all occurances of the same chunk contents. If not, then you can choose to only add a chunk once. Also, it is possible but very unlikely that different chunk contents could occur that have the same hash value, so you would get a collision on add but with different contents; that can be ignored because it is so unlikely.

Next we scan through the "new" file. For each chunk (that's not a zero-run chunk) we take the hash value and look it up in the hash table. If found, this gives us a very likely chunk match. I don't like the idea of just trusting the hash value, I want to make patches that I know 100% of the time are valid, so I then verify that the actual contents of those chunks are the same with a memcmp.

Something that you can optionally do before looking up each chunk hash is to see if the previous chunk match can be extended to cover the current chunk. Say you're in a region that is unchanged; you might have 1 MB or more of contiguous bytes that match the previous contents, but the CDC chunk cutter has still cut that into 512 or 1024 byte chunks, depending on your target chunk length. Rather than look up all those chunks in the hash one by one, once you find the first one, you can just extend it across the whole shared data run. This does not have a significant effect on speed in my experience, but can help to reduce fragmentation of the match locations.

So, we now have all the chunks of the "new" file, and each tracks a location where it matches in the "previous" file, if any. But so far these matches are all at the CDC hash-determined split points. The big improvement we can do is to extend those matches when possible.

Any time there is a repeated portion of data, we are likely to get CDC chunk boundaries somewhere inside that. In order for a chunk to match, it must first have boundaries at the same data-determined place, and to find those boundaries we use a rolling hash with a 64-byte window, so you generally will only start matching a region after at least 64 bytes where the new/old are equal. eg :

D = different
S = same

  ][          ][           ][            ][    CDC chunk boundaries

chunks with any D different bytes won't match
in the SS run, the CDC boundaries ][ would be in the same place
  only after enough same bytes get into the rolling hash window to forget all D bytes
so that chunk will find a match

  ][ no match ][MMMMMMMMMMM][ no match   ][

initial patch only matches the middle portion :


grow region while bytes match :

            <- MMMMMMMMMMMMM ->

You can fail to match the first and last target_chunk_length on any repeated portion, so we can gain a lot by expanding the match to include those.

To do that we just take any match region that borders a no-match region and see if we can step that boundary by matching more bytes. You keep stepping the boundary into the no match region, possibly until it disappears completely. (for efficiency it's important not to do something like a std::vector erase when a chunk is reduced to zero length; you can just treat zero-length regions as being "deleted" and clean them up in one pass at the end).

You can take adjacent "no match" regions and merge them into a single larger region (again being careful about how you delete). Another clean up step which you may want to do is to take adjacent match regions and see if they can be merged. If they point to adjacent areas of the source file, then they can be trivially merged into a longer match without checking any bytes. (ie. if the first chunk matches from pos P in the source, and the next chunk matches at pos (P+length_of_preceding_chunk), they can be trivially merged). If they point to different areas, then you need to check to see if the neighboring chunk also occurs at that different area.

Note that this requires both the "new" and "previous" files to be completely in memory so that we can do data matches in them, so rsync couldn't do this because it is designed to work only incrementally on one file at a time. You could extend the rdiff patch maker to do something similar to this (growing matches beyond chunks) if you made it keep both files in memory. Similarly you can do the CDC patch scheme more like rsync, and just trust the hashes without verifying the data match, and not grow the match regions beyond the CDC split points. For my use, I want an offline patcher that makes the minimum possible patch size (given the chunk size limitation, etc.), so I prefer to require the old and new file in memory and do these extra steps.

So we now have a set of chunks that are either {zero run, matched, no match}. We can output a patch :

For all chunks :
write chunk length
write trinary type of chunk : {zero run, matched, no match}
if chunk is a zero run , done
if chunk is a match, write match-from location
if chunk is a non-match, write "length" bytes of payload data

Okay, that's our patch generater. Let's compare the worst case big-O complexity to the rsync method, and also look at where we spend time in practice.

On a file of length N

We scan over all the bytes looking for zero runs or CDC boundaries
CDC scan steps a rolling hash byte by byte, comparing against threshold and tracking min

CDC scan should typically stop after ~ chunk_len bytes, but can go as far as 4*chunk_len before we enforce a limit,
but we then go back to the location of the min value, which can be less than chunk_len bytes
So worst case we actually scan 16X over each byte.  This is very unlikely.  In practice 2X does occur.
This is O(N) in any case.

Once we have chunks, we hash them all with XXH3 or similar; this is O(N)

We do this same scan on both the previous & new file (unlike rsync, it's symmetric).

For target chunk len L, we make (N/L) chunks typically.
We add (N/L) hashes from the previous file to a hash table.  This is O(N/L).

On the new file we go over the (N/L) chunks and look each up in the hash table.  This is O(N/L).
The hash table lookup tends to be a cache miss as the hash table is unlikely to fit in cache.
We then verify the hash match is an actual byte match.  This is L bytes per chunk over (N/L) chunks, so O(N).

We then try to extend matched chunks if they border unmatched chunks.  This is L bytes per chunk over (N/L) chunks, so O(N). 
Try to merge neighboring matching chunks.  I think this is O(N*log(N)), but rarely takes any time in practice.

Essentially all the steps are O(N) and there are no steps that can have terrible degeneracies that make us much slower. The worst spot where unfavorable data can make us slower is in the CDC hash boundary finding step. If we are consistently hitting the fragment length limit without finding a natural hash-determined cut point, and then being sent back to the location of the "min", that does cause a significant increase in run time (maybe 2X slower). As long as the data is making nice random rolling hash values, that is statistically unlikely, but on degenerate data that has patterns which put you into cycles of the rolling hash, it does occur.

The speeds in practice are :

Make "signature" for both files :  1.7 cycles/byte
  find zero runs and CDC roll hash boundaries : 1.5 cycles/byte
  hash chunks with XXH3 : 0.2 cycles/byte

Make patches from signatures : 0.5 cycles/byte  (around 500 cycles per chunk)
  add to hash table and look up hashes : 300 cycles/chunk
  verify bytes match and output patch record : 200 cycles/chunk

The signature phase is done on both the "previous" and "new" file, so 2N bytes
The "make patches" phase is done only on the "new" file, so N bytes

(not counting file IO). Obviously a major issue for speed in the real world will be file IO and parallelism, which we will address in the next part.

Aside on things that I currently do NOT do :

When you send a chunk in the patch that was not matched, you are sending the bytes in that chunk. I do not attempt to compress those bytes. As noted back in part 1 the problem I am considering is "coarse grain patching" , where the data I am working on I assume to be already compressor and/or encrypted in chunks, so that the raw bytes are not compressible. If that is not the case, then there are a variety of options for further compressing the "no match" raw bytes. (perhaps the optimal way would be to find the portion of the source file that they are most similar to, but don't exactly match, and send them as a "fine grain" delta from that region; this is large problem space to explore).

I currently only consider patching the "new" file from the "previous" file (or multiple previous files in a file set). You could certainly also patch against preceding data in the new file, or from other files in the "new" set. Because the CDC "signature" is symmetric, you compute the same thing on the new and old files, the same kind of matching you do to find equal chunks in the previous set could be used to find matches of chunks within your own file or against other files in the new set.

I currently assume that the working set fits in memory. eg. on a typical 256 GB machine we can patch roughly 110 GB of "previous" data to 110 GB of "new" data (leaving some room for hash tables and patch output). If you need to be able to generate patches on data sets larger than memory, that can still be done efficiently, but adds complication and is not addressed here.

old rants