RFR: 8259886 : Improve SSL session cache performance and scalability

classic Classic list List threaded Threaded
10 messages Options
Reply | Threaded
Open this post in threaded view
|

RFR: 8259886 : Improve SSL session cache performance and scalability

djelinski
Under certain load, MemoryCache operations take a substantial fraction of the time needed to complete SSL handshakes. This series of patches improves performance characteristics of MemoryCache, at the cost of a functional change: expired entries are no longer guaranteed to be removed before live ones. Unused entries are still removed before used ones, and cache performance no longer depends on its capacity.

First patch in the series contains a benchmark that can be run with `make test TEST="micro:CacheBench"`.
Baseline results before any MemoryCache changes:
Benchmark       (size)  (timeout)  Mode  Cnt     Score    Error  Units
CacheBench.put   20480      86400  avgt   25    83.653 ?  6.269  us/op
CacheBench.put   20480          0  avgt   25     0.107 ?  0.001  us/op
CacheBench.put  204800      86400  avgt   25  2057.781 ? 35.942  us/op
CacheBench.put  204800          0  avgt   25     0.108 ?  0.001  us/op
there's a nonlinear performance drop between 20480 and 204800 entries, probably attributable to CPU cache thrashing. Beyond 204800 entries the cache scales more linearly.

Benchmark results after the 2nd and 3rd patches are pretty similar, so I'll only copy one:
Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
CacheBench.put   20480      86400  avgt   25  0.146 ? 0.002  us/op
CacheBench.put   20480          0  avgt   25  0.108 ? 0.002  us/op
CacheBench.put  204800      86400  avgt   25  0.150 ? 0.001  us/op
CacheBench.put  204800          0  avgt   25  0.106 ? 0.001  us/op
The third patch improves worst-case times on a mostly idle cache by scattering removal of expired entries over multiple `put` calls. It does not affect performance of an overloaded cache.

The 4th patch removes all code that clears cached values before handing them over to the GC. [This comment](https://github.com/openjdk/jdk/commit/5859a0320334bfb6b46b62eb16b4c387641f4a2a#diff-c6bd583a97fbc4f471621fee7eab37c63718cdb6932ce357fa403cfda4b32b6fL346) stated that clearing values was supposed to be a GC performance optimization. It wasn't. Benchmark results after that commit:
Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
CacheBench.put   20480      86400  avgt   25  0.113 ? 0.001  us/op
CacheBench.put   20480          0  avgt   25  0.075 ? 0.002  us/op
CacheBench.put  204800      86400  avgt   25  0.116 ? 0.001  us/op
CacheBench.put  204800          0  avgt   25  0.072 ? 0.001  us/op
I wasn't expecting that much of an improvement, and don't know how to explain it.

The 40ns difference between cache with and without a timeout can be attributed to 2 `System.currentTimeMillis()` calls; they were pretty slow on my VM.

-------------

Commit messages:
 - Do not invalidate objects before GC
 - Always expunge on put
 - Stop scanning expired entries after first non-expired one
 - Add cache microbenchmark

Changes: https://git.openjdk.java.net/jdk/pull/2255/files
 Webrev: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=00
  Issue: https://bugs.openjdk.java.net/browse/JDK-8259886
  Stats: 138 lines in 3 files changed: 85 ins; 40 del; 13 mod
  Patch: https://git.openjdk.java.net/jdk/pull/2255.diff
  Fetch: git fetch https://git.openjdk.java.net/jdk pull/2255/head:pull/2255

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v2]

djelinski
On Mon, 1 Feb 2021 18:35:56 GMT, djelinski <[hidden email]> wrote:

>> Hm, maybe you just misunderstand how this makefile construct works. If you just want to add "--add-exports java.base/sun.security.util=ALL-UNNAMED", then that's all you should put in this assignment.
>
> yeah, I'm new to makefiles. Let me try that...

Removed. I could have sworn I tried this... but apparently I didn't. Thanks for the suggestion!

-------------

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v2]

djelinski
In reply to this post by djelinski
On Tue, 2 Feb 2021 02:37:39 GMT, Xue-Lei Andrew Fan <[hidden email]> wrote:

> the benchmark performance improvement is a trade-off between CPU and memory, by keeping expired entries while putting a new entry in the cache

Not exactly. The memory use is capped by cache size. The patch is a trade off between the cache's hit/miss ratio and CPU; we will get faster cache access at the cost of more frequent cache misses.

All calls to `put()` remove expired items from the front of the queue, and never perform a full scan. `get()` calls shuffle the queue, moving the accessed item to the back. Compare this to original code where `put()` only removed expired items when the cache overflowed, and scanned the entire cache.
Let me give some examples.
**Example 1**: insertions at a fast pace leading to cache overflows and no expirations. Here the new implementation improves performance. Consider a cache with size=4, timeout=10, and the following sequence of events:
T=1, put(1);
T=2, put(2);
T=3, put(3);
T=4, put(4);
Cache contents after these calls (same in old and new scenario). Queue order: least recently accessed items on the left, most recently accessed on the right. _K_ denotes cache key, _exp_ denotes entry expiration time and is equal to insertion time _T_ plus timeout:

|K=1, exp=11|K=2, exp=12|K=3, exp=13|K=4, exp=14|

If we now add another item to the queue, it will overflow. Here's where the implementations behave differently, but the outcome is identical: old one scans the entire list for expired entries; new one improves performance by ending the search for expired entries after encountering the first non-expired entry (which is the first entry in the above example). The end result is the same in both cases - oldest (least recently accessed) item is dropped:
T=5, put(5)

|K=2, exp=12|K=3, exp=13|K=4, exp=14|K=5, exp=15|

**Example 2**: insertions at a moderate pace, with interleaved reads. Here the new implementation improves performance, but at a possible cost of wasting cache capacity on expired entries. Consider a cache with size=4, timeout=7, and the following sequence of events:
T=1, put(1);
T=3, put(3);
T=5, put(5);
T=7, put(7);
T=7, get(1);
Cache contents after these calls:

|K=3, exp=10|K=5, exp=12|K=7, exp=14|K=1, exp=8|

`get(1)` operation moved item with K=1 to the back of the queue.

If we wait for item with K=1 to expire and then add another item to the queue, it will overflow. Here's where the implementations behave differently, and the outcome is different: old one scans the entire list for expired entries, finds entry with K=1 and drops it; new one gives up after first non-expired entry (which is the first entry), and drops the first entry.

So, when we perform:
T=9, put(9);

Old implementation will get:
|K=3, exp=10|K=5, exp=12|K=7, exp=14|K=9, exp=16|

New implementation will get:
|K=5, exp=12|K=7, exp=14|K=1, exp=8(expired)|K=9, exp=16|

Note that:
- an attempt to retrieve expired item (i.e. `get(1)`) will immediately remove that item from cache, making room for other items
- retrieving a non-expired item will move it to the back of the queue, behind all expired items

**Example 3**: insertions at a slow pace, where most items expire before queue overflows. Here the new implementation improves memory consumption. Consider a cache with size=4, timeout=1, and the following sequence of events:
T=1, put(1);
T=3, put(3);
T=5, put(5);
T=7, put(7);
Every cache item is expired at then point when a new one is added. Old implementation only removes expired entries when cache overflows, so all entries will still be there:

|K=1, exp=2(expired)|K=3, exp=4(expired)|K=5, exp=6(expired)|K=7, exp=8|

New implementation removes expired entries on every put, so after the last put only one entry is in the cache:

|K=7, exp=8|

After another put the old implementation will encounter a cache overflow and remove all expired items.

Let me know if that helped.

> add two more types of benchmark: get the entries and remove the entries

Both these operations are constant-time, both before and after my changes. Do you expect to see some oddities here, or do we just want a benchmark that could be used to compare other implementations?

> increase the size to some big scales, like 2M and 20M

Can do. Do you think it makes sense to also benchmark the scenario where GC kicks in and collects soft references?

> it may change the behavior of a few JDK components

Of all uses of Cache, only `SSLSessionContextImpl` (TLS session cache), `StatusResponseManager` (OCSP stapling) and `LDAPCertStoreImpl` (I'm not familiar with that one) set expiration timeout; when the timeout is not set, the behavior is exactly the same as before.
`StatusResponseManager` is constantly querying the same keys, and is liberally sized, so I don't expect much of an impact.
TLS session cache changes may result in fewer session resumptions and more full handshakes; I expect the cache performance improvement to more than offset the CPU cycles lost on full handshakes.

How do I file a CSR?

Also, what do you think about the changes done in `Do not invalidate objects before GC` 5859a03 commit? They offer a minor performance improvement, but if clearing the values before GC is an important security feature of this cache, I'm prepared to drop that commit.

-------------

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v3]

djelinski
In reply to this post by djelinski
> Under certain load, MemoryCache operations take a substantial fraction of the time needed to complete SSL handshakes. This series of patches improves performance characteristics of MemoryCache, at the cost of a functional change: expired entries are no longer guaranteed to be removed before live ones. Unused entries are still removed before used ones, and cache performance no longer depends on its capacity.
>
> First patch in the series contains a benchmark that can be run with `make test TEST="micro:CacheBench"`.
> Baseline results before any MemoryCache changes:
> Benchmark       (size)  (timeout)  Mode  Cnt     Score    Error  Units
> CacheBench.put   20480      86400  avgt   25    83.653 ?  6.269  us/op
> CacheBench.put   20480          0  avgt   25     0.107 ?  0.001  us/op
> CacheBench.put  204800      86400  avgt   25  2057.781 ? 35.942  us/op
> CacheBench.put  204800          0  avgt   25     0.108 ?  0.001  us/op
> there's a nonlinear performance drop between 20480 and 204800 entries, probably attributable to CPU cache thrashing. Beyond 204800 entries the cache scales more linearly.
>
> Benchmark results after the 2nd and 3rd patches are pretty similar, so I'll only copy one:
> Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
> CacheBench.put   20480      86400  avgt   25  0.146 ? 0.002  us/op
> CacheBench.put   20480          0  avgt   25  0.108 ? 0.002  us/op
> CacheBench.put  204800      86400  avgt   25  0.150 ? 0.001  us/op
> CacheBench.put  204800          0  avgt   25  0.106 ? 0.001  us/op
> The third patch improves worst-case times on a mostly idle cache by scattering removal of expired entries over multiple `put` calls. It does not affect performance of an overloaded cache.
>
> The 4th patch removes all code that clears cached values before handing them over to the GC. [This comment](https://github.com/openjdk/jdk/commit/5859a0320334bfb6b46b62eb16b4c387641f4a2a#diff-c6bd583a97fbc4f471621fee7eab37c63718cdb6932ce357fa403cfda4b32b6fL346) stated that clearing values was supposed to be a GC performance optimization. It wasn't. Benchmark results after that commit:
> Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
> CacheBench.put   20480      86400  avgt   25  0.113 ? 0.001  us/op
> CacheBench.put   20480          0  avgt   25  0.075 ? 0.002  us/op
> CacheBench.put  204800      86400  avgt   25  0.116 ? 0.001  us/op
> CacheBench.put  204800          0  avgt   25  0.072 ? 0.001  us/op
> I wasn't expecting that much of an improvement, and don't know how to explain it.
>
> The 40ns difference between cache with and without a timeout can be attributed to 2 `System.currentTimeMillis()` calls; they were pretty slow on my VM.

djelinski has updated the pull request incrementally with one additional commit since the last revision:

  Add benchmarks for get and remove

-------------

Changes:
  - all: https://git.openjdk.java.net/jdk/pull/2255/files
  - new: https://git.openjdk.java.net/jdk/pull/2255/files/34949970..abe0e238

Webrevs:
 - full: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=02
 - incr: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=01-02

  Stats: 76 lines in 1 file changed: 54 ins; 4 del; 18 mod
  Patch: https://git.openjdk.java.net/jdk/pull/2255.diff
  Fetch: git fetch https://git.openjdk.java.net/jdk pull/2255/head:pull/2255

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v2]

djelinski
In reply to this post by djelinski
On Tue, 2 Feb 2021 12:19:22 GMT, djelinski <[hidden email]> wrote:

>> If I get the patch right, the benchmark performance improvement is a trade-off between CPU and memory, by keeping expired entries while putting a new entry in the cache.  I'm not very sure of the performance impact on memory and GC collections.  Would you mind add two more types of benchmark: get the entries and remove the entries, for cases that there are 1/10, 1/4, 1/3 and 1/2 expired entries in the cache?  And increase the size to some big scales, like 2M and 20M.
>>
>> It looks like a spec update as it may change the behavior of a few JDK components (TLS session cache, OCSP stapling response cache, cert store cache, certificate factory, etc), because of "expired entries are no longer guaranteed to be removed before live ones".  I'm not very sure of the impact. I may suggest to file a CSR and have more eyes to check the compatibility impact before moving forward.
>
>> the benchmark performance improvement is a trade-off between CPU and memory, by keeping expired entries while putting a new entry in the cache
>
> Not exactly. The memory use is capped by cache size. The patch is a trade off between the cache's hit/miss ratio and CPU; we will get faster cache access at the cost of more frequent cache misses.
>
> All calls to `put()` remove expired items from the front of the queue, and never perform a full scan. `get()` calls shuffle the queue, moving the accessed item to the back. Compare this to original code where `put()` only removed expired items when the cache overflowed, and scanned the entire cache.
> Let me give some examples.
> **Example 1**: insertions at a fast pace leading to cache overflows and no expirations. Here the new implementation improves performance. Consider a cache with size=4, timeout=10, and the following sequence of events:
> T=1, put(1);
> T=2, put(2);
> T=3, put(3);
> T=4, put(4);
> Cache contents after these calls (same in old and new scenario). Queue order: least recently accessed items on the left, most recently accessed on the right. _K_ denotes cache key, _exp_ denotes entry expiration time and is equal to insertion time _T_ plus timeout:
>
> |K=1, exp=11|K=2, exp=12|K=3, exp=13|K=4, exp=14|
>
> If we now add another item to the queue, it will overflow. Here's where the implementations behave differently, but the outcome is identical: old one scans the entire list for expired entries; new one improves performance by ending the search for expired entries after encountering the first non-expired entry (which is the first entry in the above example). The end result is the same in both cases - oldest (least recently accessed) item is dropped:
> T=5, put(5)
>
> |K=2, exp=12|K=3, exp=13|K=4, exp=14|K=5, exp=15|
>
> **Example 2**: insertions at a moderate pace, with interleaved reads. Here the new implementation improves performance, but at a possible cost of wasting cache capacity on expired entries. Consider a cache with size=4, timeout=7, and the following sequence of events:
> T=1, put(1);
> T=3, put(3);
> T=5, put(5);
> T=7, put(7);
> T=7, get(1);
> Cache contents after these calls:
>
> |K=3, exp=10|K=5, exp=12|K=7, exp=14|K=1, exp=8|
>
> `get(1)` operation moved item with K=1 to the back of the queue.
>
> If we wait for item with K=1 to expire and then add another item to the queue, it will overflow. Here's where the implementations behave differently, and the outcome is different: old one scans the entire list for expired entries, finds entry with K=1 and drops it; new one gives up after first non-expired entry (which is the first entry), and drops the first entry.
>
> So, when we perform:
> T=9, put(9);
>
> Old implementation will get:
> |K=3, exp=10|K=5, exp=12|K=7, exp=14|K=9, exp=16|
>
> New implementation will get:
> |K=5, exp=12|K=7, exp=14|K=1, exp=8(expired)|K=9, exp=16|
>
> Note that:
> - an attempt to retrieve expired item (i.e. `get(1)`) will immediately remove that item from cache, making room for other items
> - retrieving a non-expired item will move it to the back of the queue, behind all expired items
>
> **Example 3**: insertions at a slow pace, where most items expire before queue overflows. Here the new implementation improves memory consumption. Consider a cache with size=4, timeout=1, and the following sequence of events:
> T=1, put(1);
> T=3, put(3);
> T=5, put(5);
> T=7, put(7);
> Every cache item is expired at then point when a new one is added. Old implementation only removes expired entries when cache overflows, so all entries will still be there:
>
> |K=1, exp=2(expired)|K=3, exp=4(expired)|K=5, exp=6(expired)|K=7, exp=8|
>
> New implementation removes expired entries on every put, so after the last put only one entry is in the cache:
>
> |K=7, exp=8|
>
> After another put the old implementation will encounter a cache overflow and remove all expired items.
>
> Let me know if that helped.
>
>> add two more types of benchmark: get the entries and remove the entries
>
> Both these operations are constant-time, both before and after my changes. Do you expect to see some oddities here, or do we just want a benchmark that could be used to compare other implementations?
>
>> increase the size to some big scales, like 2M and 20M
>
> Can do. Do you think it makes sense to also benchmark the scenario where GC kicks in and collects soft references?
>
>> it may change the behavior of a few JDK components
>
> Of all uses of Cache, only `SSLSessionContextImpl` (TLS session cache), `StatusResponseManager` (OCSP stapling) and `LDAPCertStoreImpl` (I'm not familiar with that one) set expiration timeout; when the timeout is not set, the behavior is exactly the same as before.
> `StatusResponseManager` is constantly querying the same keys, and is liberally sized, so I don't expect much of an impact.
> TLS session cache changes may result in fewer session resumptions and more full handshakes; I expect the cache performance improvement to more than offset the CPU cycles lost on full handshakes.
>
> How do I file a CSR?
>
> Also, what do you think about the changes done in `Do not invalidate objects before GC` 5859a03 commit? They offer a minor performance improvement, but if clearing the values before GC is an important security feature of this cache, I'm prepared to drop that commit.

Added benchmarks for get & remove. Added tests for 5M cache size. Switched time units to nanoseconds. Results:
Benchmark           (size)  (timeout)  Mode  Cnt    Score   Error  Units
CacheBench.get       20480      86400  avgt   25   62.999 ? 2.017  ns/op
CacheBench.get       20480          0  avgt   25   41.519 ? 1.113  ns/op
CacheBench.get      204800      86400  avgt   25   67.995 ? 4.530  ns/op
CacheBench.get      204800          0  avgt   25   46.439 ? 2.222  ns/op
CacheBench.get     5120000      86400  avgt   25   72.516 ? 0.759  ns/op
CacheBench.get     5120000          0  avgt   25   53.471 ? 0.491  ns/op
CacheBench.put       20480      86400  avgt   25  117.117 ? 3.424  ns/op
CacheBench.put       20480          0  avgt   25   73.582 ? 1.484  ns/op
CacheBench.put      204800      86400  avgt   25  116.983 ? 0.743  ns/op
CacheBench.put      204800          0  avgt   25   73.945 ? 0.515  ns/op
CacheBench.put     5120000      86400  avgt   25  230.878 ? 7.582  ns/op
CacheBench.put     5120000          0  avgt   25  192.526 ? 7.048  ns/op
CacheBench.remove    20480      86400  avgt   25   39.048 ? 2.036  ns/op
CacheBench.remove    20480          0  avgt   25   36.293 ? 0.281  ns/op
CacheBench.remove   204800      86400  avgt   25   43.899 ? 0.895  ns/op
CacheBench.remove   204800          0  avgt   25   43.046 ? 0.759  ns/op
CacheBench.remove  5120000      86400  avgt   25   51.896 ? 0.640  ns/op
CacheBench.remove  5120000          0  avgt   25   51.537 ? 0.536  ns/op

-------------

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v2]

djelinski
On Thu, 4 Feb 2021 19:36:24 GMT, Xue-Lei Andrew Fan <[hidden email]> wrote:

>> Added benchmarks for get & remove. Added tests for 5M cache size. Switched time units to nanoseconds. Results:
>> Benchmark           (size)  (timeout)  Mode  Cnt    Score   Error  Units
>> CacheBench.get       20480      86400  avgt   25   62.999 ? 2.017  ns/op
>> CacheBench.get       20480          0  avgt   25   41.519 ? 1.113  ns/op
>> CacheBench.get      204800      86400  avgt   25   67.995 ? 4.530  ns/op
>> CacheBench.get      204800          0  avgt   25   46.439 ? 2.222  ns/op
>> CacheBench.get     5120000      86400  avgt   25   72.516 ? 0.759  ns/op
>> CacheBench.get     5120000          0  avgt   25   53.471 ? 0.491  ns/op
>> CacheBench.put       20480      86400  avgt   25  117.117 ? 3.424  ns/op
>> CacheBench.put       20480          0  avgt   25   73.582 ? 1.484  ns/op
>> CacheBench.put      204800      86400  avgt   25  116.983 ? 0.743  ns/op
>> CacheBench.put      204800          0  avgt   25   73.945 ? 0.515  ns/op
>> CacheBench.put     5120000      86400  avgt   25  230.878 ? 7.582  ns/op
>> CacheBench.put     5120000          0  avgt   25  192.526 ? 7.048  ns/op
>> CacheBench.remove    20480      86400  avgt   25   39.048 ? 2.036  ns/op
>> CacheBench.remove    20480          0  avgt   25   36.293 ? 0.281  ns/op
>> CacheBench.remove   204800      86400  avgt   25   43.899 ? 0.895  ns/op
>> CacheBench.remove   204800          0  avgt   25   43.046 ? 0.759  ns/op
>> CacheBench.remove  5120000      86400  avgt   25   51.896 ? 0.640  ns/op
>> CacheBench.remove  5120000          0  avgt   25   51.537 ? 0.536  ns/op
>
> Thank you for the comment. The big picture is more clear to me now.
>
>> Example 2:
>> Old implementation will get:
>> |K=3, exp=10|K=5, exp=12|K=7, exp=14|K=9, exp=16|
>>
>> New implementation will get:
>> |K=5, exp=12|K=7, exp=14|K=1, exp=8(expired)|K=9, exp=16|
>
> K=3 is not expired yet, but get removed, while K=1 is kept.  This behavior change may cause more overall performance hurt than improving the cache put/get performance.  For example, it need to grab the value remotely.   A full handshake or OCSP status grabbing could counteract all the performance gain with the cache update.
>
>> All calls to put() remove expired items from the front of the queue, and never perform a full scan. get() calls shuffle the queue, moving the accessed item to the back. Compare this to original code where put() only removed expired items when the cache overflowed, and scanned the entire cache.
>
> I think the idea that put() remove expired items from the front of the queue is good.  I was wondering if it is an option to have the get() method that removed expired items until the 1st un-expired item, without scan the full queue and change the order of the queue.  But there is still an issue that the SoftReference may have clear an item, which may be still valid.
>
> In general, I think the get() performance is more important than put() method, as get() is called more frequently.  So we should try to keep the cache small if possible.
>
>>> increase the size to some big scales, like 2M and 20M
>>
>> Can do. Do you think it makes sense to also benchmark the scenario where GC kicks in and collects soft references?
>
> In the update, the SoftReference.clear() get removed.  I'm not sure of the impact of the enqueued objects any longer.  In theory, it could improve the memory use, which could counteract the performance gain in some situation.
>
>> Also, what do you think about the changes done in Do not invalidate objects before GC 5859a03 commit?
> See above, it is a concern to me that the soft reference cannot be cleared with this update.
>
>> How do I file a CSR?
> Could you edit the bug: https://bugs.openjdk.java.net/browse/JDK-8259886?  In the more drop down menu, there is a "Create CSR" option.  You can do it if we have an agreement about the solution and impact.

Thanks for your review! Some comments below.
> A full handshake or OCSP status grabbing could counteract all the performance gain with the cache update.

Yes, but that's unlikely. Note that K=3 is before K=1 in the queue only because 3 wasn't used since 1 was last used. This means that either K=3 is used less frequently than K=1, or that all cached items are in active use. In the former case we don't lose much by dropping K=3, and in the latter we are dealing with full cache at all times, which means that most `put()`s remove un-expired items anyway.
> get() [..] without [..] change the order of the queue

If we do that, frequently used entries will be evicted at the same age as never used ones. This means we will have to recompute (full handshake/fresh OCSP) both the frequently used and the infrequently used entries. It's better to recompute only the infrequently used ones, and reuse the frequently used as long as possible - we will do less work that way.
> get() performance is more important [..] so we should try to keep the cache small if possible

I don't see the link; could you explain?
> In the update, the SoftReference.clear() get removed. I'm not sure of the impact of the enqueued objects any longer. In theory, it could improve the memory use, which could counteract the performance gain in some situation.

That's the best part: no objects ever get enqueued! We only called `clear()` right before losing the last reference to `SoftCacheEntry` (which is the `SoftReference`). When GC collects the `SoftReference`, it does not enqueue anything. GC only enqueues the `SoftReference` when it collects the referenced object (session / OCSP response) without collecting the `SoftReference` (cache entry) itself.
I wasn't sure about this at first, so I added code to crash the benchmark if `emptyQueue` ever finds anything. It didn't crash.
> Could you edit the bug

I'd need an account on the bug tracker first.

-------------

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v2]

djelinski
On Wed, 10 Feb 2021 00:44:57 GMT, Xue-Lei Andrew Fan <[hidden email]> wrote:

> I may think it differently. It may be hard to know the future frequency of an cached item based on the past behaviors. For the above case, I'm not sure that K=3 is used less frequently than K=1. Maybe, next few seconds, K=1 could be more frequently.

I agree that such prediction might not be 100% accurate. But, quick google search reveals that there are [many](https://www.usenix.org/system/files/hotstorage20_paper_eytan.pdf) [articles](https://link.springer.com/article/10.1007/PL00009255) that claim that LRU caches offer better hit rates than FIFO, especially for in-memory caches.
> I would like a solution to following the timeout specification: keep the newer items if possible.

That's a trivial change; all we need to do is change `true` to `false` [here](https://github.com/openjdk/jdk/blob/abe0e238bd25adb1ddd2b655613899bfa063cd85/src/java.base/share/classes/sun/security/util/Cache.java#L268). But, as stated above, LRU is better than FIFO, so I wouldn't want to do that.

I could keep LRU and add another linked list that would store items in the order of their expiration dates; then we could quickly scan that list for expired items. Note: the order of expiration dates is not necessarily the order of insertion, because 1) `System.currentTimeMillis()` is not monotonic - it can move back when something changes the system time, 2) the expiration date is calculated at insertion time, so if someone changes the timeout on a non-empty cache, new items may have shorter expiration time than old ones. So, I'd either need to address that first (change `currentTimeMillis` to `nanoTime` and store creation time instead of expiration time), or use insertion sort for adding items (which would get very slow if either of the above mentioned situations happened).
Let me know your thoughts.

-------------

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v4]

djelinski
In reply to this post by djelinski
> Under certain load, MemoryCache operations take a substantial fraction of the time needed to complete SSL handshakes. This series of patches improves performance characteristics of MemoryCache, at the cost of a functional change: expired entries are no longer guaranteed to be removed before live ones. Unused entries are still removed before used ones, and cache performance no longer depends on its capacity.
>
> First patch in the series contains a benchmark that can be run with `make test TEST="micro:CacheBench"`.
> Baseline results before any MemoryCache changes:
> Benchmark       (size)  (timeout)  Mode  Cnt     Score    Error  Units
> CacheBench.put   20480      86400  avgt   25    83.653 ?  6.269  us/op
> CacheBench.put   20480          0  avgt   25     0.107 ?  0.001  us/op
> CacheBench.put  204800      86400  avgt   25  2057.781 ? 35.942  us/op
> CacheBench.put  204800          0  avgt   25     0.108 ?  0.001  us/op
> there's a nonlinear performance drop between 20480 and 204800 entries, probably attributable to CPU cache thrashing. Beyond 204800 entries the cache scales more linearly.
>
> Benchmark results after the 2nd and 3rd patches are pretty similar, so I'll only copy one:
> Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
> CacheBench.put   20480      86400  avgt   25  0.146 ? 0.002  us/op
> CacheBench.put   20480          0  avgt   25  0.108 ? 0.002  us/op
> CacheBench.put  204800      86400  avgt   25  0.150 ? 0.001  us/op
> CacheBench.put  204800          0  avgt   25  0.106 ? 0.001  us/op
> The third patch improves worst-case times on a mostly idle cache by scattering removal of expired entries over multiple `put` calls. It does not affect performance of an overloaded cache.
>
> The 4th patch removes all code that clears cached values before handing them over to the GC. [This comment](https://github.com/openjdk/jdk/commit/5859a0320334bfb6b46b62eb16b4c387641f4a2a#diff-c6bd583a97fbc4f471621fee7eab37c63718cdb6932ce357fa403cfda4b32b6fL346) stated that clearing values was supposed to be a GC performance optimization. It wasn't. Benchmark results after that commit:
> Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
> CacheBench.put   20480      86400  avgt   25  0.113 ? 0.001  us/op
> CacheBench.put   20480          0  avgt   25  0.075 ? 0.002  us/op
> CacheBench.put  204800      86400  avgt   25  0.116 ? 0.001  us/op
> CacheBench.put  204800          0  avgt   25  0.072 ? 0.001  us/op
> I wasn't expecting that much of an improvement, and don't know how to explain it.
>
> The 40ns difference between cache with and without a timeout can be attributed to 2 `System.currentTimeMillis()` calls; they were pretty slow on my VM.

djelinski has updated the pull request incrementally with two additional commits since the last revision:

 - Switch cache to FIFO order
 - Order of expiration should match insertion

-------------

Changes:
  - all: https://git.openjdk.java.net/jdk/pull/2255/files
  - new: https://git.openjdk.java.net/jdk/pull/2255/files/abe0e238..c7b064f0

Webrevs:
 - full: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=03
 - incr: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=02-03

  Stats: 33 lines in 1 file changed: 1 ins; 3 del; 29 mod
  Patch: https://git.openjdk.java.net/jdk/pull/2255.diff
  Fetch: git fetch https://git.openjdk.java.net/jdk pull/2255/head:pull/2255

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v5]

djelinski
In reply to this post by djelinski
> Under certain load, MemoryCache operations take a substantial fraction of the time needed to complete SSL handshakes. This series of patches improves performance characteristics of MemoryCache, at the cost of a functional change: expired entries are no longer guaranteed to be removed before live ones. Unused entries are still removed before used ones, and cache performance no longer depends on its capacity.
>
> First patch in the series contains a benchmark that can be run with `make test TEST="micro:CacheBench"`.
> Baseline results before any MemoryCache changes:
> Benchmark       (size)  (timeout)  Mode  Cnt     Score    Error  Units
> CacheBench.put   20480      86400  avgt   25    83.653 ?  6.269  us/op
> CacheBench.put   20480          0  avgt   25     0.107 ?  0.001  us/op
> CacheBench.put  204800      86400  avgt   25  2057.781 ? 35.942  us/op
> CacheBench.put  204800          0  avgt   25     0.108 ?  0.001  us/op
> there's a nonlinear performance drop between 20480 and 204800 entries, probably attributable to CPU cache thrashing. Beyond 204800 entries the cache scales more linearly.
>
> Benchmark results after the 2nd and 3rd patches are pretty similar, so I'll only copy one:
> Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
> CacheBench.put   20480      86400  avgt   25  0.146 ? 0.002  us/op
> CacheBench.put   20480          0  avgt   25  0.108 ? 0.002  us/op
> CacheBench.put  204800      86400  avgt   25  0.150 ? 0.001  us/op
> CacheBench.put  204800          0  avgt   25  0.106 ? 0.001  us/op
> The third patch improves worst-case times on a mostly idle cache by scattering removal of expired entries over multiple `put` calls. It does not affect performance of an overloaded cache.
>
> The 4th patch removes all code that clears cached values before handing them over to the GC. [This comment](https://github.com/openjdk/jdk/commit/5859a0320334bfb6b46b62eb16b4c387641f4a2a#diff-c6bd583a97fbc4f471621fee7eab37c63718cdb6932ce357fa403cfda4b32b6fL346) stated that clearing values was supposed to be a GC performance optimization. It wasn't. Benchmark results after that commit:
> Benchmark       (size)  (timeout)  Mode  Cnt  Score   Error  Units
> CacheBench.put   20480      86400  avgt   25  0.113 ? 0.001  us/op
> CacheBench.put   20480          0  avgt   25  0.075 ? 0.002  us/op
> CacheBench.put  204800      86400  avgt   25  0.116 ? 0.001  us/op
> CacheBench.put  204800          0  avgt   25  0.072 ? 0.001  us/op
> I wasn't expecting that much of an improvement, and don't know how to explain it.
>
> The 40ns difference between cache with and without a timeout can be attributed to 2 `System.currentTimeMillis()` calls; they were pretty slow on my VM.

djelinski has updated the pull request incrementally with two additional commits since the last revision:

 - Avoid unproductive cache scans
 - Revert Cache changes

-------------

Changes:
  - all: https://git.openjdk.java.net/jdk/pull/2255/files
  - new: https://git.openjdk.java.net/jdk/pull/2255/files/c7b064f0..f9bc386a

Webrevs:
 - full: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=04
 - incr: https://webrevs.openjdk.java.net/?repo=jdk&pr=2255&range=03-04

  Stats: 96 lines in 1 file changed: 59 ins; 3 del; 34 mod
  Patch: https://git.openjdk.java.net/jdk/pull/2255.diff
  Fetch: git fetch https://git.openjdk.java.net/jdk pull/2255/head:pull/2255

PR: https://git.openjdk.java.net/jdk/pull/2255
Reply | Threaded
Open this post in threaded view
|

Re: RFR: 8259886 : Improve SSL session cache performance and scalability [v2]

djelinski
In reply to this post by djelinski
On Mon, 22 Feb 2021 21:31:21 GMT, Xue-Lei Andrew Fan <[hidden email]> wrote:

>> Actually there's a much easier solution to reduce the number of slow `put()`s without making any behavioral changes.
>> The cache object could store the earliest expire time, and then exit `expungeExpiredEntries()` early when current time is earlier than the earliest expire time - when it is, we know that there are no expired items in the queue and we can skip the scan entirely.
>> @XueleiFan do you think the above is worth exploring?
>
>> Actually there's a much easier solution to reduce the number of slow `put()`s without making any behavioral changes.
>> The cache object could store the earliest expire time, and then exit `expungeExpiredEntries()` early when current time is earlier than the earliest expire time - when it is, we know that there are no expired items in the queue and we can skip the scan entirely.
>> @XueleiFan do you think the above is worth exploring?
>
> Definitely, I think it is a good improvement.  Actually, it is a surprise to me that the current code is not working this way.
>
> Sorry, I was/am on vacation, and the review could be delayed for a few days.

I reverted all earlier Cache changes, and added a new commit that caches the earliest expire time of all cached items. The observable behavior of the new code is identical to original - items are removed from cache at exactly the same time as before; we only skip scanning the cache when we know that there are no expired items inside.

The performance is substantially improved. There can be at most `cache size` scans in every `cache lifetime` period, which is roughly one scan every 4 seconds with the default SSL session cache settings. This is much better than possibly scanning on every `put()` that was possible before the changes.

My reduced set of benchmarks produced the following values:
Benchmark       (size)  (timeout)  Mode  Cnt    Score   Error  Units
CacheBench.put   20480      86400  avgt   25  148.345 ? 1.970  ns/op
CacheBench.put   20480          0  avgt   25  108.598 ? 3.787  ns/op
CacheBench.put  204800      86400  avgt   25  151.318 ? 1.872  ns/op
CacheBench.put  204800          0  avgt   25  106.650 ? 1.080  ns/op
which is comparable to what was observed with the previous commits.

-------------

PR: https://git.openjdk.java.net/jdk/pull/2255