Re: RFR: 8225239: Refactor NetworkInterface lookups
Claes, This is great work and long overdue. Thanks for taking care of it. - Michael. On 04/07/2019, 20:20, Claes Redestad wrote: Hi, please review this patch that refactors native java.net.NetworkInterface lookup logic in a few ways to address both pre-existing and recent regressions: Bug:https://bugs.openjdk.java.net/browse/JDK-8225239 Webrev: http://cr.openjdk.java.net/~redestad/8225239/open.01/ - adds a package-private method isBoundInetAddress that shortcuts some of the paths taken by existing methods when all we want to know is whether an IP address is bound to some interface on this system. This allows us to iterate over only IPv4 interfaces if the sought after IP address is an IPv4 address and vice versa for IPv6 addresses. This means a small (1.1x) to large (5x) speedup on a variety of systems, depending on how many IPv4 vs IPv6 interfaces exist on the system.. - refactors how the Windows-specific GetIPAddrTable API is used: current implementation looks up the global IP address table over and over, once per bound address, which means time spent by operations like NetworkInterface.getByInetAddress grows at least quadratically with the number of interfaces on a system. Refactoring these methods so that GetIPAddrTable is called once means a dramatic improvement on some of the API calls, e.g getByInetAddress cost drops from ~100ms to ~10ms on one of our larger test systems. - Added some additional tests and a microbenchmark. Testing: tier1-4 (some still ongoing, but all green so far) Thanks! /Claes
Re: RFR: 8225239: Refactor NetworkInterface lookups
Hi Claes, That looks like a good cleanup and improvement. Thanks for having looked into this! best regards, -- daniel On 04/07/2019 21:20, Claes Redestad wrote: Hi, please review this patch that refactors native java.net.NetworkInterface lookup logic in a few ways to address both pre-existing and recent regressions: Bug: https://bugs.openjdk.java.net/browse/JDK-8225239 Webrev: http://cr.openjdk.java.net/~redestad/8225239/open.01/ - adds a package-private method isBoundInetAddress that shortcuts some of the paths taken by existing methods when all we want to know is whether an IP address is bound to some interface on this system. This allows us to iterate over only IPv4 interfaces if the sought after IP address is an IPv4 address and vice versa for IPv6 addresses. This means a small (1.1x) to large (5x) speedup on a variety of systems, depending on how many IPv4 vs IPv6 interfaces exist on the system.. - refactors how the Windows-specific GetIPAddrTable API is used: current implementation looks up the global IP address table over and over, once per bound address, which means time spent by operations like NetworkInterface.getByInetAddress grows at least quadratically with the number of interfaces on a system. Refactoring these methods so that GetIPAddrTable is called once means a dramatic improvement on some of the API calls, e.g getByInetAddress cost drops from ~100ms to ~10ms on one of our larger test systems. - Added some additional tests and a microbenchmark. Testing: tier1-4 (some still ongoing, but all green so far) Thanks! /Claes
Re: RFR: 8225239: Refactor NetworkInterface lookups
Thanks, Michael! /Claes On 2019-07-05 10:02, Michael McMahon wrote: Claes, This is great work and long overdue. Thanks for taking care of it. - Michael. On 04/07/2019, 20:20, Claes Redestad wrote: Hi, please review this patch that refactors native java.net.NetworkInterface lookup logic in a few ways to address both pre-existing and recent regressions: Bug: https://bugs.openjdk.java.net/browse/JDK-8225239 Webrev: http://cr.openjdk.java.net/~redestad/8225239/open.01/ - adds a package-private method isBoundInetAddress that shortcuts some of the paths taken by existing methods when all we want to know is whether an IP address is bound to some interface on this system. This allows us to iterate over only IPv4 interfaces if the sought after IP address is an IPv4 address and vice versa for IPv6 addresses. This means a small (1.1x) to large (5x) speedup on a variety of systems, depending on how many IPv4 vs IPv6 interfaces exist on the system.. - refactors how the Windows-specific GetIPAddrTable API is used: current implementation looks up the global IP address table over and over, once per bound address, which means time spent by operations like NetworkInterface.getByInetAddress grows at least quadratically with the number of interfaces on a system. Refactoring these methods so that GetIPAddrTable is called once means a dramatic improvement on some of the API calls, e.g getByInetAddress cost drops from ~100ms to ~10ms on one of our larger test systems. - Added some additional tests and a microbenchmark. Testing: tier1-4 (some still ongoing, but all green so far) Thanks! /Claes
Re: RFR: 8225239: Refactor NetworkInterface lookups
Thanks, Daniel! /Claes On 2019-07-05 12:07, Daniel Fuchs wrote: Hi Claes, That looks like a good cleanup and improvement. Thanks for having looked into this! best regards, -- daniel On 04/07/2019 21:20, Claes Redestad wrote: Hi, please review this patch that refactors native java.net.NetworkInterface lookup logic in a few ways to address both pre-existing and recent regressions: Bug: https://bugs.openjdk.java.net/browse/JDK-8225239 Webrev: http://cr.openjdk.java.net/~redestad/8225239/open.01/ - adds a package-private method isBoundInetAddress that shortcuts some of the paths taken by existing methods when all we want to know is whether an IP address is bound to some interface on this system. This allows us to iterate over only IPv4 interfaces if the sought after IP address is an IPv4 address and vice versa for IPv6 addresses. This means a small (1.1x) to large (5x) speedup on a variety of systems, depending on how many IPv4 vs IPv6 interfaces exist on the system.. - refactors how the Windows-specific GetIPAddrTable API is used: current implementation looks up the global IP address table over and over, once per bound address, which means time spent by operations like NetworkInterface.getByInetAddress grows at least quadratically with the number of interfaces on a system. Refactoring these methods so that GetIPAddrTable is called once means a dramatic improvement on some of the API calls, e.g getByInetAddress cost drops from ~100ms to ~10ms on one of our larger test systems. - Added some additional tests and a microbenchmark. Testing: tier1-4 (some still ongoing, but all green so far) Thanks! /Claes
