Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

bypass PageCache for InMemoryLayer::get_values_reconstruct_data #8183

Closed
3 of 4 tasks
Tracked by #7386
problame opened this issue Jun 27, 2024 · 4 comments
Closed
3 of 4 tasks
Tracked by #7386

bypass PageCache for InMemoryLayer::get_values_reconstruct_data #8183

problame opened this issue Jun 27, 2024 · 4 comments
Assignees

Comments

@problame
Copy link
Contributor

problame commented Jun 27, 2024

part of epic #7386

bit of prior discussion in https://neondb.slack.com/archives/C033RQ5SPDH/p1719411245662839


InMemoryLayer::get_values_reconstruct_data uses read_blob, which internally uses the PageCache for block access.

Switch it to vectored reads that bypass the PageCache.

However, we want to deliver equivalent performance compared to the current code in the case where the current code, in one call, reads multiple blobs from the same 8kb EphemeralFile page.

Strategy for this (planned together with @VladLazar ):

  1. store the blob lengths in the in-memory btree
  • avoid consuming more memory space by using u32 instead of u64 for offset. u32 is enough if we cap EphemeralFile to 4GiB, which is way larger than we want it to go anyways 3.
  1. Get rid of the whole blob_io business for InMemoryLayer, we don't need it if we store offset and length in the in-memory index.
  2. For get_values_reconstruct_data, feed the (offset, length) pairs directly into the VectoredReadBuilder (after sorting them in offset order, so the builder can merge adjacent blob reads as needed)

Tasks

@problame problame self-assigned this Jun 27, 2024
@problame problame changed the title eliminate read-path PageCach'ing of InMemoryLayer blocks bypass PageCache for InMemoryLayer::get_values_reconstruct_data Jun 27, 2024
problame added a commit that referenced this issue Jul 2, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
VladLazar pushed a commit that referenced this issue Jul 8, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
VladLazar pushed a commit that referenced this issue Jul 8, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
VladLazar pushed a commit that referenced this issue Jul 8, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
VladLazar pushed a commit that referenced this issue Jul 8, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
VladLazar pushed a commit that referenced this issue Jul 8, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
VladLazar pushed a commit that referenced this issue Jul 8, 2024
part of #7418

# Motivation

(reproducing #7418)

When we do an `InMemoryLayer::write_to_disk`, there is a tremendous
amount of random read I/O, as deltas from the ephemeral file (written in
LSN order) are written out to the delta layer in key order.

In benchmarks (#7409) we can
see that this delta layer writing phase is substantially more expensive
than the initial ingest of data, and that within the delta layer write a
significant amount of the CPU time is spent traversing the page cache.

# High-Level Changes

Add a new mode for L0 flush that works as follows:

* Read the full ephemeral file into memory -- layers are much smaller
than total memory, so this is afforable
* Do all the random reads directly from this in memory buffer instead of
using blob IO/page cache/disk reads.
* Add a semaphore to limit how many timelines may concurrently do this
(limit peak memory).
* Make the semaphore configurable via PS config.

# Implementation Details

The new `BlobReaderRef::Slice` is a temporary hack until we can ditch
`blob_io` for `InMemoryLayer` => Plan for this is laid out in
#8183

# Correctness

The correctness of this change is quite obvious to me: we do what we did
before (`blob_io`) but read from memory instead of going to disk.

The highest bug potential is in doing owned-buffers IO. I refactored the
API a bit in preliminary PR
#8186 to make it less
error-prone, but still, careful review is requested.

# Performance

I manually measured single-client ingest performance from `pgbench -i
...`.

Full report:
https://neondatabase.notion.site/2024-06-28-benchmarking-l0-flush-performance-e98cff3807f94cb38f2054d8c818fe84?pvs=4

tl;dr:

* no speed improvements during ingest,  but
* significantly lower pressure on PS PageCache (eviction rate drops to
1/3)
  * (that's why I'm working on this)
* noticable but modestly lower CPU time

This is good enough for merging this PR because the changes require
opt-in.

We'll do more testing in staging & pre-prod.

# Stability / Monitoring

**memory consumption**: there's no _hard_ limit on max `InMemoryLayer`
size (aka "checkpoint distance") , hence there's no hard limit on the
memory allocation we do for flushing. In practice, we a) [log a
warning](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L5741-L5743)
when we flush oversized layers, so we'd know which tenant is to blame
and b) if we were to put a hard limit in place, we would have to decide
what to do if there is an InMemoryLayer that exceeds the limit.
It seems like a better option to guarantee a max size for frozen layer,
dependent on `checkpoint_distance`. Then limit concurrency based on
that.

**metrics**: we do have the
[flush_time_histo](https://github.com/neondatabase/neon/blob/23827c6b0d400cbb9a972d4d05d49834816c40d1/pageserver/src/tenant/timeline.rs#L3725-L3726),
but that includes the wait time for the semaphore. We could add a
separate metric for the time spent after acquiring the semaphore, so one
can infer the wait time. Seems unnecessary at this point, though.
@problame
Copy link
Contributor Author

This week:

Next week:

  • perf qualification in pre-prod
    • Can we do this earlier, e.g., by deploying a main build to pre-prod during this week?
  • global rollout

@koivunej
Copy link
Member

Additional extraordinary deploy to eu-west-1 before global rollout.

@problame
Copy link
Contributor Author

problame commented Sep 2, 2024

Results from pre-prod:

  • extraordinary deploy to eu-west-1 happened on Friday Aug 30 ~18:00 UTC, right before the prod-like cloudbench start
  • => no WARN or ERR in the logs
  • => getpage latencies unchanged
  • => pageserver page cache records no more InMemoryLayer accesses (see screenshot below)
  • no measurable impact on overall page cache access or miss rate (expected)
  • WalReceiverConnectionHandler and MgmtRequest and InitialLogicalSizeCalculation benefitted but they're all low traffic

Image

Image

@problame
Copy link
Contributor Author

problame commented Sep 5, 2024

production rollout => no more InMemoryLayer pages in PS PageCache (dashboard)

Image

No significant impact on overall PS PageCache performance due to the small role that InMemoryLayer plays generally in terms of access rate.

The improvements in PS PageCache performance are due to #8184 which rolled out to ap-southeast-* this week.

See this thread for details: https://neondb.slack.com/archives/C033RQ5SPDH/p1725530975830579

@problame problame closed this as completed Sep 5, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants