any future changes to the underlying data type for bin sizes
(such as upgrading from `uint16_t` to `uint32_t`) can be achieved
by modifying only the `cache_bin_sz_t` definition.
Signed-off-by: Xin Yang <yangxin.dev@bytedance.com>
in the dirty ecache has been limited. This patch was tested with real
workloads using ClickHouse (Clickbench Q35) on a system with 2x240 vCPUs.
The results showed a 2X in query per second (QPS) performance and
a reduction in page faults to 29% of the previous rate. Additionally,
microbenchmark testing involved 256 memory reallocations resizing
from 4KB to 16KB in one arena, which demonstrated a 5X performance
improvement.
Signed-off-by: Jiebin Sun <jiebin.sun@intel.com>
If jemalloc is linked into a shared library, the RTLD_NEXT dlsym call
may fail since RTLD_NEXT is only specified to search all objects after
the current one in the loading order, and the pthread library may be
earlier in the load order. Instead of failing immediately, attempt one
more time to find pthread_create via RTLD_GLOBAL.
Errors cascading from this were observed on FreeBSD 14.1.
Implementation inspired by idea described in "Beyond malloc efficiency
to fleet efficiency: a hugepage-aware memory allocator" paper [1].
Primary idea is to track maximum number (peak) of active pages in use
with sliding window and then use this number to decide how many dirty
pages we would like to keep.
We are trying to estimate maximum amount of active memory we'll need in
the near future. We do so by projecting future active memory demand
(based on peak active memory usage we observed in the past within
sliding window) and adding slack on top of it (an overhead is reasonable
to have in exchange of higher hugepages coverage). When peak demand
tracking is off, projection of future active memory is active memory we
are having right now.
Estimation is essentially the same as `nactive_max * (1 + dirty_mult)`.
Peak demand purging algorithm controlled by two config options. Option
`hpa_peak_demand_window_ms` controls duration of sliding window we track
maximum active memory usage in and option `hpa_dirty_mult` controls
amount of slack we are allowed to have as a percent from maximum active
memory usage. By default `hpa_peak_demand_window_ms == 0` now and we
have same behaviour (ratio based purging) that we had before this
commit.
[1]: https://storage.googleapis.com/gweb-research2023-media/pubtools/6170.pdf
Add opt.process_madvise_max_batch which determines if process_madvise is enabled
(non-zero) and the max # of regions in each batch. Added another limiting
factor which is the space to reserve on stack, which results in the max batch of
128.
Converting size to usize is what jemalloc has been done by ceiling
size to the closest size class. However, this causes lots of memory
wastes with HPA enabled. This commit changes how usize is calculated so
that the gap between two contiguous usize is no larger than a page.
Specifically, this commit includes the following changes:
1. Adding a build-time config option (--enable-limit-usize-gap) and a
runtime one (limit_usize_gap) to guard the changes.
When build-time
config is enabled, some minor CPU overhead is expected because usize
will be stored and accessed apart from index. When runtime option is
also enabled (it can only be enabled with the build-time config
enabled). a new usize calculation approach wil be employed. This new
calculation will ceil size to the closest multiple of PAGE for all sizes
larger than USIZE_GROW_SLOW_THRESHOLD instead of using the size classes.
Note when the build-time config is enabled, the runtime option is
default on.
2. Prepare tcache for size to grow by PAGE over GROUP*PAGE.
To prepare for the upcoming changes where size class grows by PAGE when
larger than NGROUP * PAGE, disable the tcache when it is larger than 2 *
NGROUP * PAGE. The threshold for tcache is set higher to prevent perf
regression as much as possible while usizes between NGROUP * PAGE and 2 *
NGROUP * PAGE happen to grow by PAGE.
3. Prepare pac and hpa psset for size to grow by PAGE over GROUP*PAGE
For PAC, to avoid having too many bins, arena bins still have the same
layout. This means some extra search is needed for a page-level request that
is not aligned with the orginal size class: it should also search the heap
before the current index since the previous heap might also be able to
have some allocations satisfying it. The same changes apply to HPA's
psset.
This search relies on the enumeration of the heap because not all allocs in
the previous heap are guaranteed to satisfy the request. To balance the
memory and CPU overhead, we currently enumerate at most a fixed number
of nodes before concluding none can satisfy the request during an
enumeration.
4. Add bytes counter to arena large stats.
To prepare for the upcoming usize changes, stats collected by
multiplying alive allocations and the bin size is no longer accurate.
Thus, add separate counters to record the bytes malloced and dalloced.
5. Change structs use when freeing to avoid using index2size for large sizes.
- Change the definition of emap_alloc_ctx_t
- Change the read of both from edata_t.
- Change the assignment and usage of emap_alloc_ctx_t.
- Change other callsites of index2size.
Note for the changes in the data structure, i.e., emap_alloc_ctx_t,
will be used when the build-time config (--enable-limit-usize-gap) is
enabled but they will store the same value as index2size(szind) if the
runtime option (opt_limit_usize_gap) is not enabled.
6. Adapt hpa to the usize changes.
Change the settings in sec to limit is usage for sizes larger than
USIZE_GROW_SLOW_THRESHOLD and modify corresponding tests.
7. Modify usize calculation and corresponding tests.
Change the sz_s2u_compute. Note sz_index2size is not always safe now
while sz_size2index still works as expected.
Arena 0 have a dedicated initialization path, which differs from
initialization path of other arenas. The main difference for the purpose
of this change is that we initialize arena 0 before we initialize
background threads. HPA shard options have `deferral_allowed` flag which
should be equal to `background_thread_enabled()` return value, but it
wasn't the case before this change, because for arena 0
`background_thread_enabled()` was initialized correctly after arena 0
initialization phase already ended.
Below is initialization sequence for arena 0 after this commit to
illustrate everything still should be initialized correctly.
* `hpa_central_init` initializes HPA Central, before we initialize every
HPA shard (including arena's 0).
* `background_thread_boot1` initializes `background_thread_enabled()`
return value.
* `pa_shard_enable_hpa` initializes arena 0 HPA shard.
```
malloc_init_hard -------------
/ / \
/ / \
/ / \
malloc_init_hard_a0_locked background_thread_boot1 pa_shard_enable_hpa
/ / \
/ / \
/ / \
arena_boot background_thread_enabled_seta hpa_shard_init
|
|
pa_central_init
|
|
hpa_central_init
```
Before this commit we had two age counters: one global in HPA central
and one local in each HPA shard. We used HPA shard counter, when we are
reused empty pageslab and HPA central counter anywhere else. They
suppose to be comparable, because we use them for allocation placement
decisions, but in reality they are not, there is no ordering guarantees
between them.
At the moment, there is no way for pageslab to migrate between HPA
shards, so we don't actually need HPA central age counter.
This adds a new autoconf flag, --disable-user-config, which disables
reading the configuration from /etc/malloc.conf or the MALLOC_CONF
environment variable. This can be useful when integrating jemalloc in a
binary that internally handles all aspects of the configuration and
shouldn't be impacted by ambient change in the environment.
The gettid() function is available on Linux in glibc only since version
2.30. There are supported distributions that still use older glibc
version. Thus add a configure check if the gettid() function is
available and extend the check in src/prof_stack_range.c so it's skipped
also when gettid() isn't available.
Fixes: https://github.com/jemalloc/jemalloc/issues/2740
When evaluating changes in HPA logic, it is useful to know internal
`hpa_shard` state. Great deal of this state is `psset`. Some of the
`psset` stats was available, but in disaggregated form, which is not
very convenient. This commit exposed `psset` counters to `mallctl`
and malloc stats dumps.
Example of how malloc stats dump will look like after the change.
HPA shard stats:
Pageslabs: 14899 (4354 huge, 10545 nonhuge)
Active pages: 6708166 (2228917 huge, 4479249 nonhuge)
Dirty pages: 233816 (331 huge, 233485 nonhuge)
Retained pages: 686306
Purge passes: 8730 (10 / sec)
Purges: 127501 (146 / sec)
Hugeifies: 4358 (5 / sec)
Dehugifies: 4 (0 / sec)
Pageslabs, active pages, dirty pages and retained pages are rows added
by this change.
Config validation was introduced at 3aae792b with main intention to fix
infinite purging loop, but it didn't actually fix the underlying
problem, just masked it. Later 47d69b4ea was merged to address the same
problem.
Options `hpa_dirty_mult` and `hpa_hugification_threshold` have different
application dimensions: `hpa_dirty_mult` applied to active memory on the
shard, but `hpa_hugification_threshold` is a threshold for single
pageslab (hugepage). It doesn't make much sense to sum them up together.
While it is true that too high value of `hpa_dirty_mult` and too low
value of `hpa_hugification_threshold` can lead to pathological
behaviour, it is true for other options as well. Poor configurations
might lead to suboptimal and sometimes completely unacceptable
behaviour and that's OK, that is exactly the reason why they are called
poor.
There are other mechanism exist to prevent extreme behaviour, when we
hugified and then immediately purged page, see
`hpa_hugify_blocked_by_ndirty` function, which exist to prevent exactly
this case.
Lastly, `hpa_dirty_mult + hpa_hugification_threshold >= 1` constraint is
too tight and prevents a lot of valid configurations.
Linux 6.1 introduced `MADV_COLLAPSE` flag to perform a best-effort
synchronous collapse of the native pages mapped by the memory range into
transparent huge pages.
Synchronous hugification might be beneficial for at least two reasons:
we are not relying on khugepaged anymore and get an instant feedback if
range wasn't hugified.
If `hpa_hugify_sync` option is on, we'll try to perform synchronously
collapse and if it wasn't successful, we'll fallback to asynchronous
behaviour.
Milliseconds are used a lot in hpa, so it is convenient to have
`nstime_ms_since` function instead of dividing to `MILLION` constantly.
For consistency renamed `nstime_msec` to `nstime_ms` as `ms` abbreviation
is used much more commonly across codebase than `msec`.
```
$ grep -Rn '_msec' include src | wc -l
2
$ grep -RPn '_ms( |,|:)' include src | wc -l
72
```
Function `nstime_msec` wasn't used anywhere in the code yet.
Option `experimental_hpa_strict_min_purge_interval` was expected to be
temporary to simplify rollout of a bugfix. Now, when bugfix rollout is
complete it is safe to remove this option.
During boot, some mutexes are not initialized yet, plus there's no point taking
many mutexes while everything is covered by the global init lock, so the locking
assumptions in some functions (e.g. background_thread_enabled_set()) can't be
enforced. Skip the lock owner check in this case.
