= New credit system design = == Peak FLOPS and efficiency == BOINC estimates the peak FLOPS of each processor. For CPUs, this is the Whetstone benchmark score. For GPUs, it's given by a manufacturer-supplied formula. Applications access memory, and the speed of a host's memory system is not reflected in its Whetstone score. So a given job might take the same amount of CPU time and a 1 GFLOPS host as on a 10 GFLOPS host. The "efficiency" of an application running on a given host is the ratio of actual FLOPS to peak FLOPS. GPUs typically have a much higher (50-100X) peak speed than GPUs. However, application efficiency is typically lower (very roughly, 10% for GPUs, 50% for CPUs). == Credit system goals == Some possible goals in designing a credit system: * Device neutrality: similar jobs should get similar credit regardless of what processor or GPU they run on. * Project neutrality: different projects should grant about the same amount of credit per day for a given host. It's easy to show that both goals can't be satisfied simultaneously when there is more than one type of processing resource. == The first credit system == In the first iteration of BOINC's credit system, "claimed credit" was defined as {{{ C1 = H.whetstone * J.cpu_time }}} There were then various schemes for taking the average or min of the claimed credit of the replicas of a job, and using that as the "granted credit". We call this system "Peak-FLOPS-based" because it's based on the CPU's peak performance. The problem with this system is that, for a given app version, efficiency can vary widely between hosts. In the above example, the 10 GFLOPS host would claim 10X as much credit, and its owner would be upset when it was granted only a tenth of that. Furthermore, the credits granted to a given host for a series of identical jobs could vary widely, depending on the host it was paired with by replication. This seemed arbitrary and unfair to users. == The second credit system == We then switched to the philosophy that credit should be proportional to number of FLOPs actually performed by the application. We added API calls to let applications report this. We call this approach "Actual-FLOPs-based". SETI@home had an application that allowed counting of FLOPs, and they adopted this system. They added a scaling factor so that the average credit per job was the same as the first credit system. Not all projects could count FLOPs, however. So SETI@home published their average credit per CPU second, and other projects continued to use benchmark-based credit, but multiplied it by a scaling factor to match SETI@home's average. This system had several problems: * It didn't address GPUs. * Project that couldn't count FLOPs still had device neutrality problems. * It didn't prevent credit cheating when single replication was used. == Goals of the new (third) credit system == * Completely automate credit - projects don't have to change code, settings, etc. * Device neutrality * Limited project neutrality: different projects should grant about the same amount of credit per CPU hour, averaged over hosts. Projects with GPU apps should grant credit in proportion to the efficiency of the apps. (This means that projects with efficient GPU apps will grant more credit on average. That's OK). == Peak FLOP Count (PFC) == This system goes back to the Peak-FLOPS-based approach, but addresses its problems in a new way. When a job is issued to a host, the scheduler specifies usage(J,D), J's usage of processing resource D: how many CPUs and how many GPUs (possibly fractional). If the job is finished in elapsed time T, we define peak_flop_count(J), or PFC(J) as {{{ PFC(J) = T * (sum over devices D (usage(J, D) * peak_flop_rate(D)) }}} Notes: * We use elapsed time instead of actual device time (e.g., CPU time). If a job uses a resource inefficiently (e.g., a CPU job that does lots of disk I/O) PFC() won't reflect this. That's OK. The key thing is that BOINC reserved the device for the job, whether or not the job used it efficiently. * usage(J,D) may not be accurate; e.g., a GPU job may take more or less CPU than the scheduler thinks it will. Eventually we may switch to a scheme where the client dynamically determines the CPU usage. For now, though, we'll just use the scheduler's estimate. The idea of the system is that granted credit for a job J is proportional to PFC(J), but is normalized in the following ways: == Cross-version normalization == If a given application has multiple versions (e.g., CPU and GPU versions) the average granted credit is the same for each version. The adjustment is always downwards: we maintain the average PFC*(V) of PFC() for each app version, find the minimum X, then scale each app version's jobs by (X/PFC*(V)). The result is called "Version-Normalized Peak FLOP Count", or VNPFC(J). Notes: * This mechanism provides device neutrality. * This addresses the common situation where an app's GPU version is much less efficient than the CPU version (i.e. the ratio of actual FLOPs to peak FLOPs is much less). To a certain extent, this mechanism shifts the system towards the "Actual FLOPs" philosophy, since credit is granted based on the most efficient app version. It's not exactly "Actual FLOPs", since the most efficient version may not be 100% efficient. * Averages are computed as a moving average, so that the system will respond quickly as job sizes change or new app versions are deployed. == Cross-project normalization == If an application has both CPU and GPU versions, then the version normalization mechanism uses the CPU version as a "sanity check" to limit the credit granted for GPU jobs. Suppose a project has an app with only a GPU version, so there's no CPU version to act as a sanity check. If we grant credit based only on GPU peak speed, the project will grant much more credit per GPU hour than other projects, violating limited project neutrality. The solution to this is: if an app has only GPU versions, then we scale its granted credit by the average scaling factor for that GPU type among projects that do have both CPU and GPU versions. This factor is obtained from a central BOINC server. Notes: * Projects will run a periodic script to update the scaling factors. * Rather than GPU type, we'll actually use plan class, since e.g. the average efficiency of CUDA 2.3 apps may be different from that of CUDA 2.1 apps. * Initially we'll obtain scaling factors from large projects that have both GPU and CPU apps (e.g., SETI@home). Eventually we'll use an average (weighted by work done) over multiple projects. == Host normalization == For a given application, all hosts should get the same average granted credit per job. To ensure this, for each application A we maintain the average VNPFC*(A), and for each host H we maintain VNPFC*(H, A). The "claimed credit" for a given job J is then {{{ VNPFC(J) * (VNPFC*(A)/VNPFC*(H, A)) }}} Notes: * VNPFC* is averaged over jobs, not hosts. * Both averages are exponential recent averages, so that they respond to changes in job sizes and app versions characteristics. * This assumes that all hosts are sent the same distribution of jobs. There are two situations where this is not the case: a) job-size matching, and b) GPUGrid.net's scheme for sending some (presumably larger) jobs to GPUs with more processors. To deal with this, we can weight jobs by workunit.rsc_flops_est. == Replication and cheating == Host normalization mostly eliminates the incentive to cheat by claiming excessive credit (i.e., by falsifying benchmark scores or elapsed time). An exaggerated claim will increase VNPFC*(H,A), causing subsequent claimed credit to be scaled down proportionately. This means that no special cheat-prevention scheme is needed for single replications; granted credit = claimed credit. For jobs that are replicated, granted credit is be set to the min of the valid results (min is used instead of average to remove the incentive for cherry-picking, see below). However, there are still some possible forms of cheating. * One-time cheats (like claiming 1e304) can be prevented by capping VNPFC(J) at some multiple (say, 10) of VNPFC*(A). * Cherry-picking: suppose an application has two types of jobs, which run for 1 second and 1 hour respectively. Clients can figure out which is which, e.g. by running a job for 2 seconds and seeing if it's exited. Suppose a client systematically refuses the 1 hour jobs (e.g., by reporting a crash or never reporting them). Its VNPFC*(H, A) will quickly decrease, and soon it will be getting several thousand times more credit per actual work than other hosts! Countermeasure: whenever a job errors out, times out, or fails to validate, set the host's error rate back to the initial default, and set its VNPFC*(H, A) to VNPFC*(A) for all apps A. This puts the host to a state where several dozen of its subsequent jobs will be replicated. == Implementation == Database changes: New table "host_app_version" {{{ int host_id; int app_version_id; double avg_vnpfc; // recent average int njobs; double total_vnpfc; }}} New fields in "app_version": {{{ double avg_vnpfc; int njobs; double total_vnpfc; }}} New fields in "app": {{{ double min_avg_vnpfc; // min value of app_version.avg_vnpfc }}}