Sparse files contain large regions containing only zero bytes. Checking
that a blob only contains zeros is possible with over 100GB/s for modern
x86 CPUs. Calculating sha256 hashes is only possible with 500MB/s (or
2GB/s using hardware acceleration). Thus we can speed up the hash
calculation for all zero blobs (which always have length
chunker.MinSize) by checking for zero bytes and then using the
precomputed hash.
The all zeros check is only performed for blobs with the minimal chunk
size, and thus should add no overhead most of the time. For chunks which
are not all zero but have the minimal chunks size, the overhead will be
below 2% based on the above performance numbers.
This allows reading sparse sections of files as fast as the kernel can
return data to us. On my system using BTRFS this resulted in about
4GB/s.
Ignored packs were reported as an empty pack by EachByPack. The most
immediate effect of this is that the progress bar for rebuilding the
index reports processing more packs than actually exist.
Previously the buffer was grown incrementally inside `repo.LoadUnpacked`.
But we can do better as we already know how large the index will be.
Allocate a bit more memory to increase the chance that the buffer can be
reused in the future.
For large pack sizes we might be only interested in the first and last
blob of a pack file. Thus stream a pack file in multiple parts if the
gaps between requested blobs grow too large.
Use runtime.GOMAXPROCS(0) as worker count for CPU-bound tasks,
repo.Connections() for IO-bound task and a combination if a task can be
both. Streaming packs is treated as IO-bound as adding more worker
cannot provide a speedup.
Typical IO-bound tasks are download / uploading / deleting files.
Decoding / Encoding / Verifying are usually CPU-bound. Several tasks are
a combination of both, e.g. for combined download and decode functions.
In the latter case add both limits together. As the backends have their
own concurrency limits restic still won't download more than
repo.Connections() files in parallel, but the additional workers can
decode already downloaded data in parallel.
Use only a single not completed pack file to keep the number of open and
active pack files low. The main change here is to defer hashing the pack
file to the upload step. This prevents the pack assembly step to become
a bottleneck as the only task is now to write data to the temporary pack
file.
The tests are cleaned up to no longer reimplement packer manager
functions.
Previously, SaveAndEncrypt would assemble blobs into packs and either
return immediately if the pack is not yet full or upload the pack file
otherwise. The upload will block the current goroutine until it
finishes.
Now, the upload is done using separate goroutines. This requires changes
to the error handling. As uploads are no longer tied to a SaveAndEncrypt
call, failed uploads are signaled using an errgroup.
To count the uploaded amount of data, the pack header overhead is no
longer returned by `packer.Finalize` but rather by
`packer.HeaderOverhead`. This helper method is necessary to continue
returning the pack header overhead directly to the responsible call to
`repository.SaveBlob`. Without the method this would not be possible,
as packs are finalized asynchronously.
As MergeFinalIndex and index uploads can occur concurrently, it is
necessary for MergeFinalIndex to check whether the IDs for an index were
already set before merging it. Otherwise, we'd loose the ID of an index
which is set _after_ uploading it.
github.com/pkg/errors is no longer getting updates, because Go 1.13
went with the more flexible errors.{As,Is} function. Use those instead:
errors from pkg/errors already support the Unwrap interface used by 1.13
error handling. Also:
* check for io.EOF with a straight ==. That value should not be wrapped,
and the chunker (whose error is checked in the cases changed) does not
wrap it.
* Give custom Error methods pointer receivers, so there's no ambiguity
when type-switching since the value type will no longer implement error.
* Make restic.ErrAlreadyLocked private, and rename it to
alreadyLockedError to match the stdlib convention that error type
names end in Error.
* Same with rest.ErrIsNotExist => rest.notExistError.
* Make s3.Backend.IsAccessDenied a private function.
When given a buf that is big enough for a compressed blob but not its
decompressed contents, the copy at the end of LoadBlob would skip the
last part of the contents.
Fixes #3783.
fd05037e1a changed the allocation batch
size from 256 to 128 under the assumption that an indexEntry is 60 bytes
on amd64, but it's 64: structs are padded out to a multiple of 8 for
alignment reasons. That means we'd waste no space in malloc even without
the batch allocation, at least on 64-bit machines. While that strategy
cuts the overallocation down dramatically for many small indexes, it also
seems to slow allocation down (Go 1.18, Linux, amd64, -benchtime=2s):
name old time/op new time/op delta
DecodeIndex-8 4.67s ± 5% 4.60s ± 1% ~ (p=0.953 n=10+5)
DecodeIndexParallel-8 4.67s ± 3% 4.60s ± 1% ~ (p=0.953 n=10+5)
IndexHasUnknown-8 37.8ns ± 8% 36.5ns ±14% ~ (p=0.841 n=5+5)
IndexHasKnown-8 38.5ns ±12% 37.7ns ±10% ~ (p=0.968 n=5+5)
IndexAlloc-8 615ms ±18% 607ms ± 1% ~ (p=1.000 n=10+5)
IndexAllocParallel-8 245ms ±11% 285ms ± 6% +16.40% (p=0.001 n=10+5)
MasterIndexAlloc-8 286ms ± 9% 275ms ± 2% ~ (p=1.000 n=10+5)
LoadIndex/v1-8 27.0ms ± 4% 26.8ms ± 1% ~ (p=0.690 n=5+5)
LoadIndex/v2-8 22.4ms ± 1% 22.8ms ± 2% +1.48% (p=0.016 n=5+5)
name old alloc/op new alloc/op delta
IndexAlloc-8 446MB ± 0% 446MB ± 0% -0.00% (p=0.000 n=8+4)
IndexAllocParallel-8 446MB ± 0% 446MB ± 0% -0.00% (p=0.008 n=8+5)
MasterIndexAlloc-8 213MB ± 0% 159MB ± 0% -25.47% (p=0.000 n=10+5)
name old allocs/op new allocs/op delta
IndexAlloc-8 913k ± 0% 2632k ± 0% +188.19% (p=0.008 n=5+5)
IndexAllocParallel-8 913k ± 0% 2632k ± 0% +188.21% (p=0.008 n=5+5)
MasterIndexAlloc-8 318k ± 0% 1172k ± 0% +267.86% (p=0.008 n=5+5)
Instead, this patch sets a batch size of 4, which means no space is
wasted by malloc on 64-bit and very little on 32-bit. It still gets very
close to the savings from not allocating in batches, without requiring
special code for bits.UintSize==64. Benchmark results, again for
Linux/amd64:
name old time/op new time/op delta
DecodeIndex-8 4.67s ± 5% 4.83s ± 9% ~ (p=0.315 n=10+10)
DecodeIndexParallel-8 4.67s ± 3% 4.68s ± 4% ~ (p=0.315 n=10+10)
IndexHasUnknown-8 37.8ns ± 8% 44.5ns ±19% ~ (p=0.095 n=5+5)
IndexHasKnown-8 38.5ns ±12% 36.9ns ± 8% ~ (p=0.690 n=5+5)
IndexAlloc-8 615ms ±18% 628ms ±18% ~ (p=0.218 n=10+10)
IndexAllocParallel-8 245ms ±11% 262ms ± 9% +7.02% (p=0.043 n=10+10)
MasterIndexAlloc-8 286ms ± 9% 287ms ±13% ~ (p=1.000 n=10+10)
LoadIndex/v1-8 27.0ms ± 4% 26.8ms ± 0% ~ (p=1.000 n=5+5)
LoadIndex/v2-8 22.4ms ± 1% 22.5ms ± 0% ~ (p=0.056 n=5+5)
name old alloc/op new alloc/op delta
IndexAlloc-8 446MB ± 0% 446MB ± 0% ~ (p=1.000 n=8+10)
IndexAllocParallel-8 446MB ± 0% 446MB ± 0% -0.00% (p=0.000 n=8+8)
MasterIndexAlloc-8 213MB ± 0% 160MB ± 0% -25.02% (p=0.000 n=10+9)
name old allocs/op new allocs/op delta
IndexAlloc-8 913k ± 0% 1333k ± 0% +45.94% (p=0.000 n=8+10)
IndexAllocParallel-8 913k ± 0% 1333k ± 0% +45.94% (p=0.000 n=8+8)
MasterIndexAlloc-8 318k ± 0% 525k ± 0% +64.99% (p=0.000 n=10+10)
The allocation method indexmap.newEntry has also been rewritten in a
form that is a few instructions shorter.
This removes RunWorkers, which had become mere overhead by successive
refactors. It also ensures that each former user of that function
returns any context error that occurs, so failure to complete an
operation is always reported as an error.
